This post can also be read on our website

Before AI can take over the world, it needs a robot army. And in order to build a robot army, it needs robot factories. Which makes robotics pretty important to the future of AI.

While some AI predictions have proven hyperbolic, the question of how to make robotics more generalizable is a natural outgrowth of the rapid improvement in large language models (LLMs).

We are still a long way from robot armies. Robots today are impressive, but but only when doing a narrow task.

General purpose robotics (Gen-P) is an attempt to change that, and do for robots what LLMs did for software: move from narrow, hand-crafted behavior to systems that generalize. By “Gen-P” we mean robots that can adapt across tasks and environments with minimal task-specific programming or retraining.

In this essay, we examine how the lessons from LLMs are shaping the emerging Gen-P stack, why generalization in the physical world is fundamentally harder than in software, and the two primary paths companies are taking to build generalizable robotics systems. We then explore what this means for commercialization, investment and where durable value is most likely to emerge as robotics moves from impressive demos to systems that work reliably in the real world.

🐚 Cambrian explosion

Robotics has been an important investment category for years, but the last few have brought a sharp influx of capital into the new class of Gen-P startups.

The shift is visible in the funding numbers. Billions have flown into this space in just a few years, with more than 150% YoY growth this year alone.

Capital is concentrating around the companies seen as early leaders, including Figure AI, Skild AI, Field AI and Physical Intelligence, each of which has raised mega-rounds at multibillion-dollar valuations. At the same time, a wave of newer entrants is raising aggressively to close the gap, like Generalist, Dyna, Genesis AI and Persona AI.

🌐 The Digital Precedent: Lessons from LLMs

The LLM playbook in the digital “world of bits” is the benchmark for what investors hope to achieve in the physical “world of atoms” with Gen-P, so outlining the LLM stack and key developments is a helpful starting point.

At the foundation layer, a general-purpose model is trained by scaling data and compute.

At the application layer, that base model is adapted into an assistant using Reinforcement Learning from Human Feedback (RLHF) and other algorithms (e.g., fine-tuning, architecture choices, reasoning scaffolds) to make it more reliable and aligned.

At the distribution layer, users access the system through interfaces (apps/web) or APIs, which produce feedback (prompts, completions, ratings, outcomes) that is collected and incorporated later into new training runs—improving the next version rather than updating the deployed model in real time.

• Foundation - scale (data + compute) beat cleverness: Rich Sutton’s “Bitter Lesson“ held true. Massive compute and data (increasingly synthetic) has driven LLM performance and fueled an data center arms race to build better foundation models

• Application - algorithms have turned raw power into usable assistants: Techniques like RLHF, more efficient architectures (e.g. mixture of experts (MoE)), and “System 2” reasoning models (o1-style) transformed scaled models from chaotic engines into aligned, step-by-step thinkers. The resulting application layer models can be applied to solve a broad set of problems

• Distribution - value is shifting from models to workflows/specialization: Foundation models are becoming utilities under price and open-source pressure, while durable value accrues to specialized applications that own workflows, data and customer relationships

• Feedback - digital AI is general but not adaptive in the wild: LLMs can handle a wide range of tasks zero-shot, but once deployed they do not learn from their mistakes. User interactions don’t update the model in real time; meaningful learning still happens offline through reinforcement learning or fine-tuning or through selecting particular types of training data to improve the next generation of foundation models.

🍎 AI meets physical: The Gen-P robotics stack

The rapid evolution of LLMs deeply shapes intuitions and architectures for generalizable robotics. A Gen-P stack is emerging which is similar to that for LLMs, but with some important differences. Critically, in contrast to LLMs, which primarily understand and generate output in the form of language, generalizable robotics must understand and interact with the physical world.

At the foundation layer, a physics model is trained with data (sensor streams: vision, depth, proprioception, force/torque, etc.) and compute to predict contact, forces, object motion and the outcomes of candidate actions (i.e., a learned intuition for how the world responds).

At the application layer, that physics model is combined with hardware (HW) constraints like embodiment, sensors, actuators and end-effectors, into an application model that turns perception into motor commands for a specific robot. The application model is improved using application data (task- and environment-specific logs, demos and outcomes) plus simulation (synthetic rollouts to scale training and testing).

At the distribution layer, capability ships as a deployed robot operating in real environments, where performance and edge cases generate the next round of data for offline model updates.

Physics Model (data + compute)

Data is also the foundation of Gen-P robotics and combines with compute to create a model of the physical world. This physics model takes in sensor data such as images, depth, joint positions, etc., and predicts contact, force, object motion and the consequences of a proposed action. In effect, it provides the robot with a learned intuition about how to interact with objects: how hard to grip, how far to push, when friction will and won’t hold, and which actions are feasible given its embodiment.

Unlike language, there is no internet-scale corpus of physical interaction - this new data needs to be generated and collected, making it the core bottleneck.

There are four main approaches to collecting this data that feed into the physics model:

1. Teleoperation: humans remotely operate robots to collect demonstrations; mostly high-quality but slow, and sensitive to hardware mismatches, latency and poor force feedback

2. Simulation: synthetic data from physics engines; easy to scale and good for locomotion/coarse skills, but weak for fine manipulation, friction and deformables (e.g. soft/cloth-like objects whose shape changes)

3. Real-world demos: human videos or motion-capture rigs (e.g. body cam), plus logs from deployed robots; rich strategies and context, but often mismatched to robot embodiments

4. Direct actuation: humans wear kinematically matched end-effectors with the robot’s own sensors. Rigs can be used anywhere or directly on the production line itself, allowing data to be collected in the exact environments where robots will eventually operate. This produces diverse real-world data consistent with a robot’s action and sensing parameters

Application Model (physics model + HW)

An LLM has no body; its “body” is the command line. A robot has a specific physical form. Because of this, a physics model that can understands the physical world must still navigate the specific hardware constraints of the robot to achieve practical outcomes.

The application model is the layer through which these hardware constraints are intermediated - leveraging specific sensor input to understand the scene and actuate motors to move the robot or its end effectors to create particular outcomes.

On the level of hardware, important choices need to be made in order to allow the robot to perform certain functions. One of the most critical choices on the hardware level is that of the end-effector, the part of the robot that most directly interfaces with the physical world.

These are the key categories of end effectors, each with its own tradeoffs:

• Multi-digit: Hand-like fingers for dexterous grasping, regrasping and manipulating irregular objects and tools; these are the closest analogy to human hands and are likely to have the greatest generalizability but are also more complex and expensive

• Gripper: Mechanical pincers/jaws for robust, repeatable holding of parts across a wide range of shapes and weights

• Suction cups: Vacuum-based picking for fast handling of flat, smooth, or sealed-surface items (often in high-throughput workflows)

• Specialty handling: Specific shapes or properties (magnetic, electroadhesive, needle, hooks/forks, etc) for hard-to-grip materials and edge case manipulation; these end effectors solve fixturing and grasping problems when simple fingers or suction cups don’t work

• Tooling: Task-performing end effectors (e.g., welding, fastening, dispensing, cutting, sanding, etc.) that act on the workpiece rather than just holding it; these end effectors perform a process on the workpiece rather than just performing grasping or manipulation

Depending on the distribution strategy, the physics and application models might be quite distinct, or more tightly integrated.

Distribution: The specificity of the physical world

Some aspects of the Gen-P stack (such as the physics model) are broadly generalizable, but a crucial difference to LLMs is that robotics applications must be scaled through hardware in the physical world, with direct implications for feedback loops and go-to-market strategies.

Broadly, there are two core areas in which general robotics can be deployed: B2B applications focused on business use cases, and B2C applications aimed at consumers.

B2B applications include production-line automation (case packing, palletizing/depalletizing, kitting, quality inspection), discrete task-focused workcells (machine tending, CNC loading/unloading, welding, sorting) and delivery or handling systems in warehouses and factories. Success in B2B is defined by hard return on investment (ROI) considerations: higher uptime, lower labor cost per unit, better safety and throughput, plus deep integration into existing workflows and IT systems. Differentiation comes from tight vertical focus, fast and reliable deployment, service and support and proprietary data loops in specific environments that make the system steadily better than generic competitors.

B2C applications center on the home: mobile robots that can load and unload dishwashers, tidy and move objects, assist with cooking and laundry, monitor safety, support seniors or people with disabilities, and eventually act as a general household assistant rather than a single-purpose appliance. Success in B2C depends less on formal ROI and more on trust, daily usefulness and emotional acceptance: the robot must be safe, intuitive, affordable and reliably remove real pain points from people’s lives. Differentiation comes from doing a few high-frequency, high-annoyance jobs extremely well, building an ecosystem around the device (apps, services, integrations) and compounding advantage through in-home data and brand loyalty.

Share

🚪 The two paths to Gen-P

A crucial question sits at the core of Gen-P robotics: go wide or narrow?

There are two development paths that are being pursued in parallel. One path resembles the ambition behind LLMs. The other extends the tradition of industrial automation. Each path reflects a different assumption about how generalization will emerge in the physical world, and each implies a different data strategy, deployment strategy and business model.

1. Broad loop generalizability

The first path focuses on addressing real-world complexity upstream at the level of the physics model. The end goal is a robot that can enter a new environment and solve a task it has never explicitly seen before. This is the zero-shot dream. It leans fully into the Bitter Lesson: if you can find a reliable gradient of improvement, you can justify extraordinary investment into scaling data collection and compute until the rare-but-important failures that dominate reliability (the so-called ”long tail of edge cases”) is compressed.

In the purest version of this approach, the application model largely collapses into the physics model. Task logic, planning and control are expected to emerge from a sufficiently powerful general model rather than from hand-built task layers. The “application” becomes a thin wrapper around a broad capability engine. More pragmatic versions still retain an application layer, but treat it as temporary scaffolding. As the physics model improves, more responsibility is pulled upstream and the application layer is expected to shrink over time.

The challenge is that the feedback loops are long. Large models must be trained on huge, diverse datasets, deployed into the world, observed failing in subtle ways, and then retrained with new data. Each iteration is expensive and slow, and progress is hard to measure until systems are tested in real environments. The potential payoff is enormous, but the timelines are uncertain and the difficulty of the long tail remains largely unmapped.

2. Narrow loop specialization

The second path takes a different approach to complexity, aiming to solve for it in context. This approach focuses on the application model, using it to train a specific end effector to complete specific tasks reliably. The hardware is closely scoped to the use case.

This is the continuation of the robotics playbook that has worked until now, but it rides the wave of improvements in physics models to solve an increasingly broad set of problem. However, core to this approach is the belief that production solutions involve complexity that require use-case specific data to solve for in order to create the kind of highly reliable solution needed in production. This path has a clear business model and commercial value. It leverages shorter feedback loops, but does not aim for a zero shot solution (i.e. works in a new task/environment without additional training). It trades ambition for pragmatism and optimization in terms of hardware with an optimal bill of materials (BoM) for the application it is tied to.

The ultimate goal is to expand into a diverse set of applications with generalizable capabilities, but arrives there through concrete problem solving rather than by leveraging massive models.

🤏 Crossing the uncanny valley

The two paths to Gen-P have distinct approaches to solving the long tail of complexity in the physical world. Broad loops assume that this complexity can be primarily solved for at the model level; narrow loops assume that this needs to be solved for through concrete deployments at the application and distribution level.

LLMs have become broadly generalizable through heavy investment in foundation models. Still there are questions about whether the current architectures can continue to generalize (Ilya Sustkever, for instance, thinks that LLMs are moving back into “the age of research”).The chart contrasts a simple domain (steep drop: most probability mass in a few common cases) with a hard domain (fat tail: many low-frequency cases still matter). Robotics lives closer to the long-tail regime: after about 80-90% coverage, progress slows because reliability depends on mastering a large number of rare, messy interactions rather than just improving performance on the common cases.

The critical question for robotics is how hard is it to solve for the long tail of complexity. Evolution spent about 500 million years getting animals to reliably perceive and act in the physical world, from early nervous systems to today’s exquisitely coordinated eyes, muscles and balance. By contrast, it’s had perhaps 2-3 million years to develop cognition in hominins, and less than 200,000 years for anything like modern language. We fixate on “intelligence” because it underwrites human supremacy and because motor control feels so “basic” we barely notice it. But that’s the uncanny part: what feels trivial to us is the product of immense evolutionary investment in solving an extraordinarily hard physical control problem.

The uncanny valley is the eerie zone where an AI system looks and moves almost, but not quite, human; close enough to trigger our social expectations but just different enough that every flaw feels disturbingly obvious. Gen-P lives in a similar uncanny valley of capability. Once robots can handle 80-90% of everyday situations, the remaining long tail of rare, messy edge cases becomes glaring, and truly “solving” Gen-P means systematically climbing out of that valley by mastering the hard residue of real-world complexity. Autonomous vehicles are a great example of hard this can be.

AVs: The long road to generalizability

The problem of the long tail of complexity in solving automation problems in the physical world is highlighted by the history of autonomous vehicles. All of Elon Musk’s predictions about robotaxis were wrong, as brutally chronicled by Bloomberg.

It is relatively easy to get an autonomous vehicle to drive well for a large percentage of the time, leading to impressive demos. Google’s car project (the predecessor to Waymo) was driving competently on highways in 2009. But it has taken 16 years to transform this technology into the production version that is ready to carry passengers on highways.

It is still hard for these vehicles to have a sufficiently sophisticated world model that can deal with every rare but risky situations, especially as deployments scale rapidly.

This Waymo doesn’t “understand” why this decision is dangerous:

Solving for autonomous vehicles is different to solving for robotics. In order to have a commercial robotaxi service that operates across a city with no human driver, you need to create a vehicle that can drive on every road in the city and negotiate almost every scenario that the vehicle will encounter, including frequent interactions with other drivers and pedestrians.

But driving has relatively few inputs - you can accelerate or break and turn the steering wheel left or right. In contrast, manipulation involves moving digits across many degrees of freedom and interacting with a wide array of objects in diverse environments which are less consistently structured and regulated than a city’s roads. Adding in locomotion for humanoid robots adds even more complexity to the equation.

It’s hard to extrapolate from these differences how relatively hard these two problems are, but it seems likely that it is harder to solve for Gen-P robotics than autonomous vehicles, and the time it has taken to get AVs from a promising demonstration to a level of maturity ready to scale offers some sense of the scale of the challenge.

TCO and ROI

The main constraint in scaling autonomous vehicles is not just autonomy, it’s economics. Robotaxi fleets only work if each vehicle stays highly utilized, and if the vehicles and sensor stacks get cheap enough to justify the service.

Robotics faces the same reality. In most B2B deployments, winners will be the teams that design the robot with the right hardware for the workflow, an optimized bill of materials (BoM), and reliability that makes the payback math obvious. Because these systems are physical and must be built, shipped, installed, and serviced, total cost of ownership (TCO) and return on investment (ROI) tend to dominate what looks like a purely technical problem.

Bifurcation of capital

Investment decisions in Gen-P robotics flow as much from investors’ theories of the future as from the constraints of their own fund models. Venture capital is increasingly bifurcated between mega-funds with massive pools of capital and lower, but more predictable return expectations and smaller funds which seek to capture outsized returns but put less capital to work in achieving them.

Broad-loop generalizability (Path 1) is structurally top-heavy in capital. These companies need to raise enormous sums up front to build data collection infrastructure, train large physics models, assemble world-class research teams and carry years of burn before achieving meaningful commercial traction. That makes them natural homes for very large pools of capital that need to write $100M+ checks into stories that can plausibly support tens of billions in outcome value.

Money here is not just fuel but a weapon: it buys talent, accelerates data and compute scale, lets companies acquire promising startups and subsidizes deployments to achieve distribution.

The risk is that the improvement curve is still unclear. We don’t yet know whether Gen-P is in an “era of scaling” or still in an “era of research,” and the true TAM is conceptually enormous but operationally vague. From an IRR perspective, these bets can be spectacular if the models keep improving and one or two winners emerge, but timelines and capital intensity also create significant challenges.

Narrow-loop specialization (Path 2) has almost the opposite capital profile. These companies can’t productively ingest unlimited capital. They are constrained by how fast they can stand up software stacks, qualify hardware, integrate with customer environments and scale service teams.

The upside to narrow-loop is a far clearer line of sight to ROI: specific use cases, defined hardware BoM, measurable payback periods and repeatable sales motions. This fits better with smaller, more focused funds and growth investors looking for strong unit economics, shorter feedback loops and the potential for attractive multiples on more modest check sizes. The upside may be capped relative to a true Gen-P platform, but the IRR can be excellent if teams compound across adjacent use cases and geographies, and it may be the only viable path to long-term scale.

Crucially, this is not necessarily an either/or world. Broad-loop players will need narrow-loop deployments to generate real data, revenue and credibility, while narrow-loop winners may eventually become acquisition targets or distribution partners for the largest platforms.

Our predictions

• Durable funding: Gen-P is a huge opportunity space - the TAM is all of human physical labor - which will continue to attract significant investor interest for a sustained period of time provided there is evidence of improvement, much as we’ve seen with autonomous vehicles

• There will be multiple types of winners: Given the size of the opportunity space, there are lots of discrete problems to be solved within both the Gen-P stack and at the application layer, with the potential to build many large businesses

• TCO will matter at the application level: Application specific robots will have defensibility and earn customer lock-in if they can demonstrate reliability and optimize BoM

• Specialization: The combination of the preceding points suggests that specialization will be rewarded (as we’re starting to see with LLMs) - winners will excel in particular areas, but will also need to achieve significant scale to be durable over time

• Hype cycle: In spite of the opportunity’s scale, the realities of the physical world will prove a drag on the hyped-up demos as the long transition to systems that work reliably in context plays out, similar to AVs; along the way there will be bankruptcies, pull-backs and consolidation

• Physics models are not the obvious bet they seem to be: Much like with foundation models, significant value will flow from physics models, but competition in this space will make this a very expensive game with the risk of commoditization to the benefit of the application layer (as we’re seeing with “wrapper” companies for LLMs)

• Humanoids will be extra hard: Though the promise of humanoids is the greatest, they are trying to solve the hardest problem and will take a very long time to reach generalizability; there will be the greatest reset in this domain

• Data strategies: Certain data collection and end effector approaches will prove more effective and the market will over time consolidate in these directions; companies that determine the best approach earlier will be advantaged

• Commercialization will be the ultimate test: Venture capital will be able to bridge the gap between vision and reality, but commercial proof points will be critical to generating a durable funding advantage for winners, and real world monetization will be needed by the time that funding momentum subsides

Share

RedBlue Capital is an early stage, thesis-driven venture capital firm focused on the transformation of the real economy inc

luding transportation, logistics, manufacturing and energy infrastructure.

Subscribe now

You’re reading the RedBlue Newsletter, which offers deep takes on the intersections between mobility and technology. You can subscribe here:

Subscribe now

Let’s imagine aviation worked like car ownership. The only way to fly is in a private plane. Since flying is useful for trade and business, the government offers tax incentives for owning a plane and uses property taxes to fund airports and swathes of hanger space needed when private planes are not in use. The planes themselves are cheaper to be more affordable for individuals, but also lower quality and less reliable. To make flying more accessible, flight training is relatively straightforward; the high rate of crashes is an unfortunate but unavoidable consequence. If you need to fly but can’t afford to own a plane, you can hitch a ride with a friend - if you have the right friends.

Mapping this analogy is tricky because it seems so absurd compared to the obvious logic of today’s commercial aviation industry. Flying is cheap, safe and reliable. A mile of flying costs less than half the price of a mile of driving. That same mile is also 60x safer to fly than to drive. Sure there are important differences between cars and planes - for starters a Gulfstream costs 300x more than a Model S - but that doesn’t change the fact that the business model supporting planes has numerous advantages over vehicle ownership.

Flying is the trip economy at scale: a transportation marketplace that sells standalone rides. Purchasing trips frequently — rather than vehicles infrequently — allows for more economies of scale, more sustainable infrastructure, and more efficiency. Trip marketplaces seamlessly match buyers and sellers, and generate competition to lower costs and improve quality. On the ground, similarly structured trip marketplaces - transit, ridehailing, and delivery services - are frequently safer, greener and more efficient than those based on car ownership. This feels paradoxical since once you own a vehicle, incremental trips feel cheap, but the price of access is high and the overall system is far less efficient. Aviation is a great example of how trip marketplaces combine affordability with comfort and has lessons that are transferrable to other forms of transportation that are hiding in plane sight.

To better understand the power of the aviation trip economy in action and how it makes for a more efficient, more comfortable and more scalable system, let’s explore three aspects of this industry that are not obvious on the surface, but have magical effects when combined.

1. A Theseus’ Ship of the skies ⛵

Southwest Airlines emerged and surged to prominence in the 1970s as aviation was being deregulated. Southwest found ways to turn around planes in very short periods of time (often in less than 15 minutes) while charging dramatically less. These innovations increased both the time planes spent in the air and the utilization of seats while flying.

The Southwest model is emblematic of a broader trend in the industry: a trip marketplace pushes towards maximizing asset utilization. Planes are in use more than 40% of the time, a rate ten times higher than cars. It helps that these assets are mobile - if a particular travel corridor sees a surge or slump in passengers, routes can quickly be adjusted to better satisfy demand.

Much like the ancient Athenians maintained the Ship of Theseus, replacing each plank as it decayed, airlines are constantly maintaining, upgrading and replacing portions of a plane at the rates they are worn down - from the coat of paint on the exterior to the seats inside. This replacement cycle is a hallmark of the trip economy. It feeds back into how manufacturers build planes: at high levels of quality to endure constant use over a long period of time and modularly to allow subcomponents to be replaced and upgraded over time. And it feeds forward into how these assets are operated, financed and maintain residual value.

Jet engines are especially Thesean and the benefits are immense. Manufacturers like Rolls-Royce have developed a unique business model that couples sales with “power-by-the-hour” charges: airlines pay an hourly rate for the engine-maker to ensure the engine is functioning well. This aligns incentives: airlines want their planes to have functioning engines whenever they fly and manufacturers are compensated for ensuring this. This model also allows manufacturers to gather data and insights about how well their engines are performing, creating a positive feedback loop.

Even though planes are expensive to finance and operate, the dynamics of the flight trip marketplace squeeze out maximal utilization and ensures affordable costs to consumers.

2. Disneyland with a jet bridge 🧚

When assets are used efficiently in a trip marketplace, more consumers can benefit. But there are additional ways to increase consumer surplus. Trip marketplace offerings invite bundling: flights and airports are combined with optional premium experiences that entice the wealthy while making these services more affordable for everyone else. It might be a bit jarring to compare a terminal to a theme park; if you’ve ever slept in an airport, your experience is unlikely to compare to the Magic Kingdom. But there are some fascinating parallels.

Airports - like Disneyland - are weirdly disassociated from the physical world that surrounds them. The liminal nature of airport terminals - nestled between international borders - is the foundation of duty free. The first version, improbably, was created for the rich and famous transiting through Shannon, Ireland. Before jet engines, planes flying transatlantic needed a place to refuel and the west coast of Ireland was closer to North America than England.

The duty free segment, which sells high end products like spirits and perfume (dominated by LVMH-owned DFS) served as the blueprint for what came next. Airports have evolved into luxury malls with global brands vying for the vacant hours of the iPhone-class travelers passing through on their way to the rest of the world. Even those not spending on high end retail benefit from high quality amenities and things to do during the time they spend waiting between flights.

Dior perfume is sold alongside Turkish delight in Istanbul and Gouda cheese in Amsterdam. Airports are emblems of and gateways to the nations that host them, making them eager to help support the airport bundle. Terminals may have less charm than Disneyland but have been expanded much more aggressively. Singapore and Dubai have been built on the back of airlines centered around a hub airport, and Saudi Arabia is attempting to do something similar with Riyadh Air.

But it is airlines that are the true masters of bundling. They have their own version of Disney Dollars: Miles. Miles are the ultimate embodiment of a loyalty bundle. Miles grant access to lounges, upgrades, shorter queues, and better experiences. Airlines bundle liquor and status with their hub airports to entice travelers to pass through them. Increasingly airlines are using miles as a currency for retail within a captive terminal - such as United’s Terminal C at Newark where you can purchase food with them. And miles extend into a broader bundled ecosystem of travel and retail experiences: credit cards, hotel bookings and car rentals.

Planes have even more captive power than terminals. Seat pricing is the most powerful way in which bundles are used to capture value from those willing to spend while affording an adequate service to everyone else. Getting from point A to point B is essentially a commodity, but when traveling a long distance locked in a plane, small differences - food, comfort, a warm scented hand towel - are worth a lot. This is especially true for employees traveling for work, with an employer that is less cost sensitive than they would be if they were covering the fare themselves.

It took airlines a while to figure out price segmentation, but it allows them to offer comfort to those who can afford it and affordability to those who need it.

3. The invisible hand 🙌

A trip marketplace shines a bright light on the companies that operate within it. When big companies are responsible for large numbers of trips, policymakers get easy access to levers for better social outcomes. This has a major impact on safety. Considering how much can go wrong while flying, it’s astounding that it so rarely does. In the US last year, just 11 people died in plane crashes. In 2021, no one did. That’s in contrast to 31,850 that died in car crashes that year. What’s more, safety for the two modes are trending in opposite directions - cars are getting more dangerous and flying has been getting safer.

Flying used to be dangerous, but because every major incident is reported and investigated, lessons are learnt and systems are improved. A standardized set of guidelines for operation and maintenance makes flying planes safer - those that adhere to it have significantly lower incident rates. Meanwhile, a database that allows air and ground crews as well as air traffic controllers to report incidents anonymously helps catch problems before they escalate.

The airline industry’s obsession with safety can be a little overbearing—like when you’re told to push your bag under your window seat or forced to close your laptop during takeoff—but puts in stark contrast how lax our approach to road fatalities is.

Beyond the generally constructive attention to safety that comes with a massive, centralized marketplace, the trip economy also gives regulators powerful tools to address externalities by factoring them into pricing. I call this the trip triangle. It offers an elegant way to discourage bad outcomes and encourage good ones while creating very little friction in the system since the externalities are wrapped into individual trip prices.

Here are some of the externalities that are bundled into the price of flying, mostly invisible to consumers:

To be clear, the ease with which costs can be added to trip prices introduces the risk of mission creep - and indeed costs have grown since airlines were deregulated in the 70s. But it is still a great way to pay for new infrastructure like airport terminals. It allowed a simple way to improve security in the wake of 9/11 (even though there is a lot to criticize about the TSA and its security theater). Flying depends on airport operations like air traffic control - and these are easily built into fares. And since they are accountable for each trip, trip pricing also means that airlines can be made to pay fees when they leave us waiting or stranded.

It’s helpful to compare this with how we pay for ground transportation. Parking is bundled into the cost of real estate and paid for even by those who don’t drive. Meanwhile, the gas tax no longer pays for the infrastructure it is intended to support as vehicles have become more efficient and we transition to electric vehicles. The ownership economy makes it much harder to align incentives and externalities.

View from above 🛬

Looking down from the portal view of an airplane as it comes in to land, you can see the blinking lights of houses and office blocks, and cars moving between them.

The big difference between these two systems is business model. Aviation works within the structure of a trip economy, where consumers pay for individual journeys and a marketplace creates intense competition to improve asset utilization and customer experience while externalities are easily priced into the cost of trips. Meanwhile cars are tied to an ownership economy in which vehicles are used a tenth as much as airplanes and remain unaffordable for many Americans, locking them out of the engine of upward mobility and exacerbating inequality.

Those who can afford the convenience of a private jet can still own one - it’s just absurd to imagine the government subsidizing it. We should channel that intuition into asking whether it makes sense to channel massive subsidies into car ownership and free parking, simply because it mostly benefits the middle class instead of the rich.

Not every lesson is portable across these systems, but it is worth considering why one seems to work so much better than the other and how we can bring the benefits of the trip economy back down to earth.

Subscribe now

Want a comprehensive picture into the mobility news cycle?

Save time by reading tables we’ve created that summarize recent news, investments and events here:

Summary tables

You’re reading the RedBlue Newsletter, which offers deep takes on the intersections between mobility and technology. You can subscribe here:

Subscribe now

ChatGPT is uncannily good. It has replaced crypto, now relegated to TMZ, as the hot-topic in tech. ChatGPT has generated stratospheric levels of excitement (and apprehension) with a belief that these systems are widely generalizable.

Autonomous vehicles were once also massively hyped, but the bubble of excitement has deflated. Many are skeptical that autonomous vehicles can scale in the near term to serve many users in many cities.

The contrast in the current sentiment towards these two technologies couldn’t be more striking.

These are two axes of comparison that are helpful in considering how each technology will scale and what the consequences will be.

1. What is the safety threshold for each technology to permit wide scale deployment?

2. How complicated is the problem that’s being solved?

Safety threshold

The safety threshold for deploying autonomous vehicles is understandably high. Moving vehicles through our cities and suburbs is dangerous. They have the potential to kill people. Indeed they tragically already have. Autonomous systems need to be extremely reliable to scale.

In contrast, GPT seems to have a relatively low bar for deployment: it’s a system that generates text on a screen rather than actuating a vehicle in the physical world. We’ve been doing this kind of thing for a while and it has mostly been harmlessly inept (hey Siri!).

But though the danger appears innocuous, it might be significant. In contrast to autonomous vehicles, where the danger comes from the vehicle itself, the immediate danger with GPT comes from us. It’s really hard to predict how this could play out - as it was hard to predict how social media might lead to a Trump presidency. But it seems that it might be bad in a much more wide reaching way than car crashes are. For instance, truth murder like QAnon or Russian propaganda - which it turns out societies are quite susceptible to - could be significantly amplified if it gets industrialized with neural nets. Longer term, there are reasons to worry about power centralizing around companies like OpenAI, transparency about how these systems work and inbuilt biases in their training data, and alignment between our goals and the outputs of AI.

So we have a high bar with AV safety and a low bar that maybe should be higher with GPT. It seems likely this bar will be raised as unexpected issues start to arise.

Complexity

At first glance, driving might seem cognitively simpler than writing sonnets in Shakespearian verse. There are four directions to direct a car: forwards, backwards, left and right. In contrast, GPT outputs any word and sometimes makes up new ones. And we have had autonomous vehicles that can drive convincingly well for a lot longer than a chat interface that we can convincingly converse with.

However the task of making cars drive themselves has proven devilishly difficult - much harder than anticipated. There is a very long tail of unexpected edge cases (grannies in wheelchairs chasing ducks, people dancing on your car) that are hard to train for or appropriately respond to. Driving also requires understanding the state of mind of other road users. And a lot less of driving—the “training data”—has been digitally recorded than our writing has.

On the other hand, even though language appears to “represent a kind of pinnacle of complexity”, as Stephen Wolfram notes in a detailed exploration of GPT, it might actually be simpler than we realize. At least on the surface, ChatGPT seems to do a pretty good job.

One way to understand the difference between the two problems: AVs have to respond to the unexpected complexities of the real world; GPT only has to map a semantic system we use to model the world. That map covers a lot more things (human knowledge) but it’s been a lot better mapped out than roads in all their dynamic complexity. Language is always a reduction and the way we use it has relatively intuitive internal rules. Perhaps the shock of GPT is not so much how complex it is as revealing how unexpectedly simple we are.

So ChatGPT may be addressing a less hard problem than it would appear - but there are still many ways in which the system is making major errors. And driving is harder than expected - but still it seems like a constrained problem. We at least seem to be not far from having things that work a lot of the time in the case of Waymo and Cruise, at least when coupled with training wheels like remote control.

So autonomous vehicles have to be super safe and it’s a pretty hard problem to solve. But the fact that autonomous vehicles have proven so hard to commercialize suggests that there may be significant challenges as GPT use cases extend into more complex territory or when errors with bad consequences prove hard to quickly remedy.

Path forward

AI is to hype what igloos are to snow. Back in 2015, Sam Altman and Elon Musk discussed autonomous vehicles, claiming that they’d be here “much faster than people think”. Sam suggested they’d be here in 3-4 years. Elon put it closer to 2-3 years. According to most carmakers, we should have had autonomous vehicles roaming our cities two years ago.

It’s worth keeping this hype cycle in mind when thinking about how ChatGPT will evolve. There are three things that might happen next:

1. GPT develops and improves quickly, while AVs languish

2. They both develop quickly

3. They both slow down

Applying neural nets is alchemy: we don’t know why things that work well do so. There’s a lot of tweaking to get better outcomes. Sometimes we have unexpected rapid improvements. But when things go wrong, it’s hard to know why because the system is a black box of billions of numerical weights.

Original approaches to AVs split the driving task into perception, planning and control, but end-to-end approaches are increasingly being applied. GPT uses new things like transformers (the T in GPT). If GPT has made fundamental breakthroughs, these kinds of learnings should percolate across to AVs. So if GPT is on a rapid improvement trajectory, it seems likely that AVs shouldn’t be far behind.

But the performance step up embodied in ChatGPT might be a mirage, especially when imagining a pathway to greater sophistication and especially to AGI. Bing Chat was trained on a larger model than ChatGPT, but its performance seems worse. It’s worth remembering that although ChatGPT appears smart, it is fundamentally dumb: its engine is simply a model that predicts the best word to add next in a text output. It’s not that the improvement pathway will be linear.

A rich history of sci-fi has has seeded many expectations and fears. And it’s clear that these technologies have the potential to create massive change and generate massive value. When it works, it’s hard to distinguish from magic. When it doesn’t, it’s also hard to understand why or how to fix it. A lesson from autonomous vehicles is that we shouldn’t let sentiments get ahead of us: appearances can be uncanny.

Subscribe now

Want a comprehensive picture into the mobility news cycle?

Save time by reading tables we’ve created that summarize recent news, investments and events here:

Summary tables

This newsletter offers deep takes on the intersections between mobility and technology. You can subscribe here:

Subscribe now

America is running out of land for housing (link).

Land now accounts for 47% of U.S. home values, estimates Mr. Davis. That is up from 38% in 2012 and less than 20% in the early 1960s. The rising value of land is responsible for almost all of the surge in home values in recent decades, he said.

It seems preposterous: How could a country so rich in space have so little available for people to live on?

The answer lies at the intersection of a few compounding factors

• Rules about where & what you can build (zoning + density): A key reason the available pool of land has shrunk is that land needs to be zoned for development. The frontier of America has been closed down through bureaucratic categorization. Beyond this, what can be built in these zones has meant that often only low density housing such as single family homes are permitted. The consequence of reducing space to build and the density of what is built on it is that land markets can’t respond to growth in demand other than through price increases.

• The capacity of cities to handle growth (transportation): America is built around cars. But cars have limited throughput (link). As you increase the density of an urban space, you also increase the demands on thoroughfare, and roads quickly max out causing congestion. It’s very hard to correct for this retroactively: Once an urban environment is built, it becomes very challenging add higher throughput modes. Landowners will fight against giving up their property to create the right of way needed. But because it can’t be built, they will also fight against denser housing or new developments, since this imposes a greater burden on roadways.

This combination of regulation and car dependency has created bonsais out American cities, stunted in their growth and absurdly overpriced. This phenomena is most pronounced in coastal cities like San Francisco and LA, but it is becoming more widespread.

This article was first published in the RedBlue Newsletter on October 6, 2022.

Subscribe now

Want a comprehensive picture into the mobility news cycle?

Save time by reading tables we’ve created that summarize recent news, investments and events here:

Summary tables

This newsletter offers deep takes on the intersections between mobility and technology. You can subscribe here:

Subscribe now

Olaf was hosted by Bloomberg to discuss the Inflation Reduction Act. Below is an except from the interview. You can watch the full interview here.

The EV parts of the IRA

The IRA extends the existing $7,500 tax credit on new vehicles, while changing the rules for which vehicles and buyers qualify.

Buyers: earning cap of $150,000 per year, doubled to $300,000 for a married couple.

Vehicles: $80,000 cost cap, vehicles must be US sourced and manufactured, even down to raw materials.

In addition to the tax credit for new vehicles, there is also now a $4,000 credit for used EVs.

Tension between America First vs Climate interests

The restriction that only vehicles fully manufactured in the US qualify for the credit is a reflection of America First policies, prioritizing domestic manufacturing over other considerations such as minimizing cost or maximizing EV supply. This policy reflects a discomfort that China today is central to the EV supply chain, manufacturing the majority of battery packs, especially as tensions between the US and China spike. The policy will accelerate a shift away from Chinese manufacturing, but that shift was already happening.

In the near term, these restrictions will severely limit the number of electric vehicles that qualify for the tax credit, with around 70% of EVs expected not to qualify. This creates confusion for buyers and even manufacturers as they try to shift their strategies to adapt to the suddenly shifted regulatory landscape.

On the positive side, the new subsidy regime brings long term policy stability with a ten year time horizon.

Partial correction towards less regressive policies

These new policies are less regressive than the existing set of subsidies, which didn’t limit who or which vehicles qualified (a wealthy person buying a Hummer EV seems a suboptimal candidate for a subsidy). Furthermore, the $4,000 credit for used EVs is particularly important for less wealthy buyers.

Two questions mostly overlooked while focusing on these specifics are: (1) does this policy meaningfully move us closer to the goal of reducing emissions from transportation? (2) are subsidies to consumers buying cars the right place to direct tax dollars?

But we should be focused on electrifying miles not electrifying vehicles

You might not want to make it an either/or, but the supply of lithium and other battery materials is limited, as is the money government can put to work to achieve key goals. Directing subsidies to cars that have massive battery packs is like subsidizing sponges that will suck lithium, cobalt and nickel away from the vehicles that could best utilize these supply constrained raw materials. The metric we should focus on is electric miles, not electric vehicles.

Beyond this, we also need to find ways to reduce the amount of energy required to move people and goods: our cars are comically over-designed for the things we need them for.

We argued this point at green.red.blue when the original Build Back Better legislation was on the table.

Here are three ways this bill and the policies it introduces could be improved going forward:

• Incentivize smaller vehicles: The most tragic aspect of the IRA is that it originally included a $900 credit to purchase an ebike, but it was struck from the final version (link). It’s crazy that if you buy a new EV, you get $7,500 of the cost subsidized, but if you want to buy an electric bike, you get nothing. The relative efficiency and affordability of ebikes is effectively being punished: you could buy seven and a half RadMission ebikes for the value of the car credit (link). One way of more effectively rewarding less energy intense forms of transportation is a policy being proposed in California of a $2,500 tax credit to any adult within a household who does not own a car (link).

• Focus on fleets: Fleets - whether trucks, Amazon delivery vans, shared scooters or Citi Bike - are the vehicles with highest utilization rates and therefore the most leveraged bundle of miles to focus on decarbonizing. Beyond this, they tend to be fit for purpose and designed specifically for the use they are applied, meaning that they consume the energy they need to fulfill their purpose. Subsidies should be directed here most aggressively to accelerate adoption and manufacturing scale while bringing forward the emissions benefits of a quicker transition. There is an allowance in the IRA for commercial vehicles (not picked up by the news cycle but present in section 45W (link)). However, the subsidy cap for small commercial vehicles is equal to the baseline for passenger vehicles. Meanwhile, similar to the missing ebike subsidy, small vehicle fleets like shared scooters and delivery bikes are overlooked: It is striking how much GM benefits from the IRA (it has the most vehicles that could qualify for subsidies) even though it was bailed out in recent memory; meanwhile Bird is on its knees as city governments (including in Santa Monica where it started) tie it up with red tape and provide no support (link), even though the miles it delivers are far, far more climate friendly.

• Small vehicle infrastructure: The biggest barrier to the wide adoption of ebikes and other smaller form factors is the danger created by cars. The best way to get around this is to build viable networks of micromobility lanes. I discuss this in my book: Though Copenhagen and Seattle have a similar climate and population size, in Seattle just 3.1% of people commute by bicycle, compared with around 50% in Copenhagen. Seattle had just 3.2 miles of protected bike lanes in 2014; in contrast Copenhagen, has 513km (319 miles) of protected micromobility infrastructure in the city and an additional 167km (104 miles) of "super bike lanes" for longer distance commuters coming into the city (link).

Future bills such as the upcoming ‘22 amortization R&D tax bill do create opportunities to direct incentives towards some of these areas with greater leverage, especially an ebike credit (link).

This article was first published in the RedBlue Newsletter on August 22, 2022.

Subscribe now

Want a comprehensive picture into the mobility news cycle?

Save time by reading tables we’ve created that summarize recent news, investments and events here:

Summary tables

This newsletter offers deep takes on the intersections between mobility and technology. You can subscribe here:

Subscribe now

There’s been a lot of startup news in the transit van / commuter shuttle world. Egyptian-Emirati leader SWVL has fared better than most and recently acquired Mexican leader Urbvan for $82M (link). Pakistani breakout Airlift - which was a sort of hero of the entire country’s startup scene - folded after raising a $85m round only a few months ago (link). COVID has been the biggest problem for all of these players since commutes were their bread & butter. For the startups who have ridden that wave, there’s another existential challenge coming their way.

All this might seem like déjà vu to San Franciscans who were riding Chariot vans back in 2017, but for those who weren’t, here’s the need-to-know on why shared van startups are going to face a lot of hurdles in the next five years - and how they might overcome.

“Private” transit?

First, what defines this category? The lines blur, but most started offering corporate shuttle service and branched out into offering a broader shared vanpool network that quickly begins to look like a private transit network.

It’s no wonder that as COVID reduced commutes, these companies flung themselves into all sorts of adjacencies. Airlift picked up quick commerce, while Urbvan began running inter-city routes.

Most of the two dozen or so shared van startups around the world - including the leader SWVL - are in megacities in emerging markets. They serve the growing middle class whose households might not afford a car or want to deal with crushing traffic, need to commute across congested cities to office parks, and aspire to nicer conditions and service speeds than are available via public transit - if that transit service is available at all.

And that’s the core arbitrage. People and the companies they work for want to get around and aren’t finding adequate ways to do so. Where government investment in transport options has not kept pace with the growth of the city, that’s where these companies step in.

Conflicts between private & public

But startups serve routes whose riders can pay to make them profitable. Those might be routes already served by public transit, and this is where conflicts frequently first arise between startup and city. The startups offer a premium service to the wealthier commuters (often subsidized by their employers) who want either a nicer experience or a faster ride with fewer stops. Stripping out the few consumers who actually pay for transit hurts the public network’s financial viability.

Transport networks are networks, and the more universally they serve places, the better their service quality and economic model becomes. There are a variety of parallels one can make: consider how standard air carriers couldn’t fly you transcon for $120 without business class seats paying $3,000, or how health insurance risk pools can’t operate if young healthy people pack up and leave.

Chariot, a shared van startup which ran in San Francisco from 2014-2018, faced serious pushback from government throughout its life. SFMTA prides itself on capturing the second highest rate of commuters in public transit in the entire US - San Francisco’s goal was very much to grow its public network universality and, along with that, the relative service quality and economic health of the SFMTA. So as Chariot began to grow, SF reviewed resolutions that would, for example, outlaw Chariot from running routes that overlapped too much with public bus routes. Chariot plodded along and eventually Ford, its acquirer, shut it down about a year before the pandemic would have given them a great face-saving cover story.

Megacities in emerging markets might lack state capacity to adequately serve transport needs. And as long as there’s not a realistic hope for an adequate public network to exist, there’s less basis for governments to legitimately obstruct the growth of profitable private networks.

But the day the state decides it will grow public networks, the conflicts will begin.

Is there a future for these private shuttle companies?

Lighting a fire under government’s ass

San Francisco might be proud of having the second-highest rate of transit-based commutes in the nation, but over 65% of commuters still drive by themselves in a car.

That’s a big problem, and here’s why - cities’ lighthouse goal should be for a large portion of people to be able to readily afford quick, safe and clean ways of getting around. Cars aren’t cheap, and when everyone uses them, they gridlock a city to the point where most people can’t get anywhere easily. Shared transit, when run effectively, can comparatively get a city much closer to that lighthouse goal.

Insofar as Chariot (and Uber for that matter) may be moving people out of cars more than out of transit, there’s a long term synergistic role for some service like it in the overall shared transportation network. In Chariot’s case, it isn’t super clear what the net effect was.

One frustration we have with transit agencies’ response to companies like Chariot is its defensiveness. As agencies try to find ways to serve residents, commitment to a certain type of product - fixed line bus service as one example - may lose sight of that lighthouse goal. When it’s revealed that consumers like new products - like 15 seater mini buses that can more closely approximate door-to-door service - we think the response should be for agencies to find a way to actually co-opt and integrate these.

Is the resistance that dampened Chariot’s stride also going to hit SWVL and Urbvan? There may well be room to grow for years in their various markets before they begin to draw serious pushback from transit agencies, but in the meantime, there are other models that are taking off.

A second way: co-opt, or be co-opted by governments

As pissed off as they may be that startups are stripping their most valuable customers from public networks, governments are waking up to the fact that they need to respond to this challenge and not just block it.

First, improving standard bus services is top of mind since half the reason many people opt for alternatives has to do with route availability and scheduling. Multiple startups have hit billion-dollar valuations selling SaaS to transit agencies including Israeli players Optibus and Moovit. US startup Swiftly may not be far behind.

Second, introducing new services that look essentially like what Urbvan or SWVL do - what transit planners call demand responsive transit (DRT) - is top of many agencies’ agendas. Software and playbooks to help roll out these kinds of new products have powered software startup Via to a multi-billion dollar valuation, while Blaise Transit in Canada and Bridj in Australia are among the wave of other players on the DRT train. Jersey City started using Via in 2020 to launch a fleet of 30 vans to provide an on-demand shared service to better fill out areas of its network too sparsely populated to run normal bus routes (link). Many other cities are following suit.

In a steady-state, which business is more attractive - Via’s or SWVL’s? Well, SWVL needs to find franchisees to buy and drive cars (like Uber, they are asset-light), and in some cities it may soon find itself butting up against transit agencies that would rather have such a service either shut down or under their own control. On the flip side, SWVL makes a lot more revenue and really owns its customer and can, say, advertise to them in the vehicle or via its app. Via basically will get a much smaller cut of revenue by selling software to the city, but in some ways one can imagine it saves a lot of operational overhead by passing to the city the work of actually running the service.

An extension of this model to work with governments is to deliver transit services by winning a concession to operate on an agency’s behalf. Unlike the software model, the private player does the heavy lifting of running the service. And in contrast to the SWVL model, these companies start by partnering with or outright serving the government. This is a well established model that the French and Singaporeans are particularly good at. Established companies like RATP, TransDev, Kelois, and ComfortDelGro run bus and light rail services in cities around the world without residents realizing it isn’t actually run by their local government. Smart transit agencies like Transport for London make sure that they have multiple competing operators bidding against one another to guarantee the best service - all the while their buses are only differentiated by small markings most people will never notice! These operators don’t really have any customer relationship, as the London payment card is managed separately by TFL.

Three of the many bus concession operators working for Transport for London - Arriva (owned by Germany’s Deutsche Bahn), Metroline (owned by Singapore’s ComfortDelGro), and the French-run RATP Group

Being one of many relatively undifferentiated operators of commodified vehicles doesn't sound like an attractive business for many startups. However, not every city has evolved to the extent London has.

India’s Chalo has a strong customer-facing element, though it operates under concession from local transit agencies

In emerging markets, there is more room for hybridization: Chalo, a startup in India that wins concessions from globally significant cities like Mumbai and Kolkata, goes far beyond being a behind-the-scenes bus driver. It takes over payments for transit and, given the relative ubiquity of transit reliance, that play may allow it to stake a claim in the high-margin, high-multiples world of payment processing for goods outside of transport. It’s grown rapidly and, almost poetically, has actually acquired India’s leading private corporate shuttle operator, Shuttl (the Chariot of India, so to speak) putting it in a position also to operate its own service for corporate customers.

There’s a broad spectrum of ways to get rides to people who want to pay for them. It’s sensible that the fastest-growing options in many emerging markets were full-stack alternatives to public transit built by companies like SWVL, Urbvan and Airlift. But now that some of these have ridden out COVID, navigating the right way to work within the broader context of urban mobility without sacrificing either the attractiveness of their business nor their license to operate is going to be the next challenge.

This article was first published in the RedBlue Newsletter on July 24, 2022.

Subscribe now

Want a comprehensive picture into the mobility news cycle?

Save time by reading tables we’ve created that summarize recent news, investments and events here:

Summary tables

This newsletter offers deep takes on the intersections between mobility and technology. You can subscribe here:

Subscribe now

While US research pioneered many of the key innovations in lithium ion technology, it was the Japanese in the ‘90s and the Koreans in the ‘00s that ran away with the market. As batteries became embedded in everything, so too did battery companies. The biggest public share offering this year was LGES, the battery division of Korea’s LG which owns roughly 22% of the global battery supply for cars (link). It was really the Chinese who took the prize in the last 2 decades, though. CATL has become one of China’s most valuable companies and supplied over 30% of global EV batteries last year (link). Incidentally, the second biggest public share offering this year after LGES was CATL, which raised nearly $7B last month (link).

European & American companies are playing catchup. Their saving grace was that they were focused on a newer, more advanced chemistry than the Chinese were scaling up. In the last two years, though, the “less sophisticated,” Chinese-mastered chemistry has proven itself a sufficient - if not better - battery for many global EV platforms. With this development, battery supply independence in the West has gone from bad to worse.

🧪 Entry level chemistries background

Two kinds of EV batteries today: Nickel & iron are the basis for the two major types of automotive lithium ion batteries. When people write LFP they mean iron, when they write NMC or NCA they mean nickel.

LFP “China’s play” (lithium iron phosphate): This is an older technology developed in the US in the ‘90s and commercialized to great success by the Japanese & Chinese for electronics & automotive. While the Japanese have largely moved on to the newer nickel chemistries, the Chinese have doubled down on iron over time. LFP is more stable, less prone to fires and broadly speaking can be abused & cycled through more aggressively with less degradation. But its energy density is fairly low at 180 to 210Wh/kg.

NMC “the next gen” (lithium nickel manganese cobalt - Tesla’s NCA variant uses aluminum): An American successor to the LFP invented in the ‘00s, the NMC was most effectively commercialized by battery makers in Korea & Japan which sell them back to American & European carmakers who believed this was the future. Nickel cell chemistries allow far higher energy density, approaching today 250Wh/kg with some line of sight up to 300Wh/kg. They are pricier than LFP per Wh and are far more finicky, requiring careful management to avoid fires or rapid degradation. They also require problematic metals - “conflict minerals” and the like - that iron batteries avoid entirely.

While iron chemistries’ improvement is reaching asymptotic limits, nickel cell advancement has been fast and “high-nickel” variants are finally stamping down on metals of concern like cobalt.

📚 How the bets were placed

Nickel batteries were a natural choice for western carmakers in the last 15 years. Nickel’s energy density enabled 200+ mile range, which is why starting with the GM Volt in 2009, every major passenger EV has launched with a nickel battery in Europe & the US.

Largely, the US & Europe are dependent on Asian manufacturers for their NMCs, though the mix of suppliers is more western-aligned “friendlies” Korea & Japan than the Chinese. The ‘09 GM Volt was powered by an NMC made by Korea’s LG, which built that battery factory in the US partly funded by Obama-era stimulus money. As nickel took over western carmakers’ agendas, so too did it become the focus for automotive lithium ion cell makers in Korea & Japan.

Meanwhile, China’s domestic market was served just fine with cheaper iron batteries, and was big enough for its major battery makers to focus mostly on building LFP technology & factories. That’s only accelerated as NMC’s cost decline has slowed, and Chinese leader BYD announced last year that it was committing entirely to LFP going forward (link).

💥 China’s upset

While non-Chinese carmakers focused almost entirely on NMC, the Chinese have nonetheless grown to account for over 40% of global EV batteries in 2021 (link) - largely by supplying nickel batteries to international customers.

In the last two years, nickel’s hold on Western car markets has come into serious question. In early 2020, estimates placed iron at between 1 and 5% of the global EV battery markets. Today, that seems to have passed 20% (link). Announcements from Renault (link), Tesla (link), Ford (link), VW (link) and many others underpin this significant shift. Given that the Chinese make essentially all LFP batteries, LFP adoption may dramatically accelerate their rise.

🤔 What changed?

A confluence of changes upset the playing field, boosting LFPs to close-enough performance to NMC, while keeping costs more attractive.

Thinking beyond the cell: The chemistry of cells isn’t alone what determines the performance of a pack, which is comprised of hundreds or thousands of these cells packaged together. Battery companies have become better at using LFP cells’ high stability to create pack designs that help close energy density gap in real-world use. Basically, LFP cells can be assembled together in packs that would catch fire if they were filled with NMC cells. With less special “packaging” needed to keep the battery safe, the effective density of an LFP battery can become closer - if not quite the same - as an NMC battery. Case in point: Tesla’s LFP Model 3 in China is advertised as getting 250 miles of range, versus a 260 mile range in the NCM model (only 4% better!). Within six months, Tesla went from zero LFP to nearly half its cars equipped with the iron batteries (link) as it has moved to selling them even in the US.

Customer needs: Part of the change comes from OEMs realizing that lower prices matter much more than slight dents in mileage for EV buyers in the lower or middle parts of the market.

Bright spots for US/Europe: Surging lithium prices impact iron batteries more than nickel ones, which can narrow the cost advantages of LFP batteries recent lithium price trends continue. Furthermore, nickel batteries are the basis for much of the cutting edge innovation - so if breakthroughs push them into 300 kWh per kilogram and above without much price jump, there could well be an NMC comeback.

This article was first published in the RedBlue Newsletter on July 5, 2022.

Subscribe now

Want a comprehensive picture into the mobility news cycle?

Save time by reading tables we’ve created that summarize recent news, investments and events here:

Summary tables

This newsletter offers deep takes on the intersections between mobility and technology. You can subscribe here:

Subscribe now

The updated offering integrates with all the screens in the vehicle (including instrument cluster) and allows deeper control of vehicle functions, including climate control.

I have seen this console - which has a lot of information like a calendar displayed - described as overwhelming in comparison to the minimalist Tesla Model 3. The critical thing is that the comparison is a natural one to make - Apple’s software allows carmakers to offer a customer experience that compares to Tesla’s. It is on digital experience that carmakers have fallen behind, while the digitized driving experience has increasingly come to define Tesla’s brand dominance (see: Is TSLA a tech stock or a fashion product?).

This critique that it is overwhelming also misses the key advantage of software: when Apple owns digital real estate, just like with the Apple Watch, it can mold according to each user’s preferences.

The question is how deep this goes. In 2016, I wrote about the automotive opportunity for Apple. Rather than building the full car, I argued that Apple can integrate deeply into vehicles leveraging its strength in silicon and UX design to capture value while avoiding the liability and pain of vehicle manufacturing:

This integration is still relatively shallow compared to where it may go. The silicon in vehicles is still very distributed and the opportunity for Apple to get deeper into the brain of a vehicle is a fascinating though complicated one.

The power of smartphone app ecosystems have seeped into consumer expectations of their cars. Tesla’s shown consumers they can expect more. But this process elsewhere has been relatively slow in comparison to China. Whereas our grandparents bought cars similar to the ones we still drive, in China, most car-owners were smartphone-owners first. They expect digital integration as a default.

So while Apple (along with Google), a smartphone company, is increasingly moving into the car, it is notable that Geely, perhaps the most important carmaker in China (which also owns Volvo and Lotus) just acquired Meizu, a once-leading smartphone company and seems interested in manufacturing phones (link). Meanwhile, it also just started launching its own array of low earth satellites with the goal of enabling low latency connectivity to its vehicles and positing to enable autonomous functionality. Finally, Geely has launched a JV with Baidu to make an “autonomous vehicle”… creatively named Jidu (as in Geely=Ji and Baidu=du) (link). The company has aggressive plans to commercialize their first vehicle, the delicately named ROBO-01 which they just unveiled (link).

Baidu is the Google of China and has been developing autonomous technology for years. But this car is not designed to be fully autonomous any time soon - autonomy here is a feature targeted at “users who like cutting-edge technologies.” Much like Tesla’s perpetually delayed FSD, the point of “autonomy” in this context is more about cabin experience and brand mystique than it is about changing the business model of transportation as Cruise or Waymo aim to do. But for now, this is a key part of the battleground in order to sell vehicles that are electric and digitized to the masses.

This article was first published in the RedBlue Newsletter on June 17, 2022.

Subscribe now

Want a comprehensive picture into the mobility news cycle?

Save time by reading tables we’ve created that summarize recent news, investments and events here:

Summary tables

This newsletter offers deep takes on the intersections between mobility and technology. You can subscribe here:

Subscribe now

As investors have switched focus from growth to profit, we’re seeing a wave of quick commerce flame-outs and layoffs.

What’s been happening?

Layoffs: Getir - 14% (link), Zapp - 10% (link), Airlift - 31% (link), Gorillas - 50% of Berlin HQ staff (link), GoPuff - already gone through two rounds of layoffs (link)

Flameouts: Jiffy (link), Buyk, Fridge No More (link) [literally]

Consolidation: Appetito acquires Lamma in MENA (link)

WTF: Zepto (link) just raised $200m and Yummy (link) raised $47m 🤔

Was there ever a point?

• Grocery isn’t as bad a business as people think: A lot of people mistakenly think grocery is defined by razor thin margins. In reality, many products (”SKUs”) have quite high gross margins. The trick is to get customers through the door to buy them - say with a luscious $5 rotisserie chicken. The first step is to exist nearby customers, and the second is to lure them in with attractive offers. Once they’re inside, they can monetize the attention of people in their stores. That last part is the key.

• Attention is the most important part of retailing: About 1/4 of CPG revenue is reinvested into advertising, a major portion of which is paid to stores - e.g. to keep your brand at eye level. When a brand launches, it has to plow hundreds of millions into getting people to pay attention and it turns out ads & paid placements in stores are more effective than Instagram.

• Being “the default” place consumers shop is therefore key - it means you are the place brands “pay to play”: Amazon and Instacart became the place many people now start their buying journeys by delivering a lot of stuff for not much money. They have enormous operational costs to do this, but it doesn’t matter - by being the starting point, they generate huge ad revenues that make them whole. In 2021 Instacart generated $550M in ad revenue, while Amazon generated a monster $31.2B.

• Quick commerce’s play to win the attention game: The proposition of “10 minute delivery” is unique, whimsical and more attention grabbing than the typical loss-leading rotisserie chicken. And remember, the question isn’t if it makes money - but that it doesn’t lose too much. What if you could sell **mostly high margin SKUs, have a lower real estate cost structure than the average corner store, and beat legacy retailers at the ads game with a purely digital storefront?

What’s the future?

If you look at the numbers (as we did in 2020-2021 looking at Gorillas and others) - well, it almost just possibly makes sense. Almost. When Walmart acquires one of these companies, this will be the reason: like the rotisserie chicken, 15 minute delivery may never need to make money. It just needs to be a shiny toy that keeps you a loyal subscriber to Walmart+ or Amazon Prime.

Meanwhile, the problem with layoffs for startups is they signal the end of growth. Startups can discount salaries through stock-options and get discounted capital on the promise that the future will be way better than the present. But that dream can only be sold when things are going up and to the right. When that stops, even the best advertising can’t help.

This article was first published in the RedBlue Newsletter on June 2, 2022.

Subscribe now

Want a comprehensive picture into the mobility news cycle?

Save time by reading tables we’ve created that summarize recent news, investments and events here:

Summary tables

This newsletter offers deep takes on the intersections between mobility and technology. You can subscribe here:

Subscribe now

Lyft’s Q1 earnings this month tanked the stock 30% in the hours after the call. The concern became contagion, dragging Uber down even before Uber’s earnings were announced. Dipping 10%, Dara rushed news to market as if to say “hold on - we’re not Lyft!” and made the rare decision to move up earnings by a day.

It worked - while Lyft remained in the gutter, Uber turned its decline around and finished off down only a few percentage points from the days prior.But why? Lyft reported a nearly $200M loss, and Uber reported a nearly $6B loss. Isn’t that worse? It certainly fed Uber naysayers

.

But why? Lyft reported a nearly $200M loss, and Uber reported a nearly $6B loss. Isn’t that worse? It certainly fed Uber naysayers.

But that $6B loss masks the real story. Let’s separate operational profit (which many tech companies call “adjusted EBITDA”) versus true profitability (generating free cash flow).

Actually, Uber is already sort of profitable

Operational profit is a question of “can Uber make more money than it spends on rides, food delivery & freight?” By successfully hitting positive adjusted EBITDA semi-regularly, Uber is now showing that this is in fact the case, and it’s how they claim “operational profitability.”

Q1 was a show of great strength for Uber’s core businesses. Food and grocery delivery kept drivers earning throughout the pandemic, priming Uber for the surge in riders we’re seeing today. CEO Dara calls this the “power of the platform.” Drivers are plentiful and getting a lot of gigs on Uber, giving Uber the breathing room to cut back on bonuses and incentives, yielding operational profitability. It is beginning to seem like Uber’s scale flywheel is spinning without help.

Understanding the $6B net income loss

On top of that operational business though, you have a lot of other heavy areas of spend. Uber previously invested a lot of resources into new market expansion (yielding a major ownership chunk of China’s Didi) and introducing new products (yielding a major chunk of autonomy leader Aurora). Recently, DiDi was slammed by China’s tech crackdown, and Aurora stock is down as well. Almost 95% of Uber’s Q1 “losses” were simply writedowns of these holdings. Those aren’t cash losses today, nor do they threaten the viability of Uber’s (increasingly profitable) operating businesses.

Back to Lyft

As Lyft didn’t have a delivery offering, its driver base shrunk a lot during COVID. Lyft had to spend heavily on incentives to get them back - and signaled it will probably need to continue doing so. That’s a big problem that triggered the stock selloff. Lyft’s marketplace flywheel is in need of more nudging. And while Uber’s mega projects (China, autonomy, flying cars) might be falling apart, its core business is a flywheel that’s ... flying (and remember that Uber’s ridehail business is much bigger AND growing top-line much faster than Lyft’s). Lyft’s picture isn’t great.

As much as it would seem that these marketplaces are commoditized for drivers and riders, Harry the Rideshare Guy told me recently that in LA Uber drivers are averaging over $30/hr, with top drivers hitting $60/hr. In contrast, Lyft drivers have lower density meaning the network has long ETAs and a lot of cancellations, making things worse for both drivers and riders.

Newspeak

OK: Operational profitability isn’t real profitability though. Ridehailing company accounting is complicated (see this for a taste) and companies have long leaned on accounting and linguistic ambiguity to mask their losses to investors. When markets were kinder, they could get away with it. But in the height of the current tempest, Dara sent an urgent email to all employees.

Breaking rank with most tech CEOs, Dara didn't wax poetic about Uber's transformative vision. He cut straight to what most big tech CEOs ignore: what investors demand. And the message was clear - not just growth, operational breakeven, "adjusted EBITDA" profitability...

The message: Uber is going to generate free cash flows.

Dara’s forecasts are finally clear: $300M in FCF this year, and $1.8B in 2023. If Q4 looks the way they are (finally) guiding, Uber won’t just be the company famous for buying growth. It will also be the first company to take ridehailing into the next stage of its life: sustainability.

This article was first published in the RedBlue Newsletter on May 19, 2022.

Subscribe now

Want a comprehensive picture into the mobility news cycle?

Save time by reading tables we’ve created that summarize recent news, investments and events here:

Summary tables

This is RedBlue Newsletter, which offers deep takes on the intersections between mobility and technology. You can subscribe here:

Subscribe now