Disclaimer: Not financial advice. Do your own research and due diligence. Please refer to this page. At the time of writing I am holding a position in UBER.

Note: The “Inference Singularity” was a beta test. Now we are hitting production mode with VOL. 01 of THE AUDIT series. Enjoy!

// SYSTEM: SILICON [02]
// DATE: JANUARY 20, 2026
// EST: 13 MIN READ

> root/executive_summary

Executive summary

The global mobility sector stands at its most significant inflection point since the advent of the smartphone. As we enter 2026, the theoretical convergence of autonomous vehicle (AV) technology and ride-hailing marketplace dynamics has transitioned into a tangible operational reality.

• The misconception: The market seems to value Uber as a cyclical transportation service facing an existential threat from Autonomous Vehicles (AVs).

• The reality: Uber is a digital utility. It is a real-time marketplace for logistics. Its moat is not the cars, but the liquidity of the network and the predictive algorithms (DeepETA, H3) that balance it.

• The thesis: AVs will not kill Uber; they will likely commoditize the hardware layer. Uber strives to become the “dispatch layer” for robotaxis, transitioning from a high-cost operator to a high-margin software monopoly.

• The “sleeper” value: Beyond the tech, Uber has quietly built a financial fortress - positive FCF, a growing “Super App” ecosystem via Uber One, the ad revenue potential and strategic equity stakes in the very AV partners meant to disrupt them.

• The risk: While the "Uber Zero" scenarios are low probability, we must monitor medium-term risks, specifically regulatory fracture, the potential for a "super app" distribution war with Google and marketplace dynamics between vertically integrated suppliers and Uber as an aggregator.

Introduction

I often read research articles on Uber. The discourse usually revolves around “Tesla vs. Waymo vs. Uber” and the fundamental misunderstanding of Uber as just an “app.”

In reality, Uber has built a moat around its platform, serving as the liquidity hub for the AV world. The bear case often sounds like this:

• It is just an app that aggregates demand and supply.

• It is a company that doesn’t own its own AV tech.

• It will face disruption as soon as platform partners achieve scale and vertical integration.

These bearish narratives ignore a fact hidden in plain sight: The company isn’t called Uber Taxis. It is called Uber Technologies Inc.

My colleague Manu at Fundamentally Sound recently published a brilliant financial dissection of the company, noting that:

“Uber’s service [is not] as simple as matching riders to drivers, [but rather solving] a complex logistical challenge. This is the value of Uber’s Marketplace Liquidity.”

I am not here to re-litigate the financials - Manu has covered the Capital layer extensively. Instead, I want to audit the Silicon. I want to take that concept of “Marketplace Liquidity” and dismantle it to see how it works mechanically. We are foregoing the balance sheet today to focus entirely on the machine.

So, how does the operating system for atoms actually work?

Fundamentally Sound

On Uber: AVs, Take Rates, and Fragmentation

One of the primary arguments against Uber as an investment can be summarized in two words: AV disruption…

Read more

2 months ago · 136 likes · 3 comments · Manu Invests

> root/silicon/audit_log_01

1. The “Taxi” Fallacy

I have a confession to make. For a long time, I thought Uber was a bad business. The narrative was easy to buy into: It’s a taxi company that owns no cars. It burns cash. It has no moat because anyone can build an app.

It is a compelling bear case. It is also wrong.

I spent the last two weeks auditing Uber - not just the 10-K, but this time their engineering stack. What I found is that Uber isn’t a transportation company. Uber is a distributed systems company building the operating system for the physical world. And I think that this is Uber’s best defense against most of these bear thesis out there.

As Chamath Palihapitiya @chamathphas argued in his seminal “Bits to Atoms” thesis, the easy gains of the consumer internet are over; the next era of value belongs to those who can bridge the gap between "Bits and Atoms".1 Uber is the first successful proof of this thesis at a global scale.

To understand the technical challenge, you have to visualize the volume. In Q3 2025 alone, Uber’s platform executed 3.5 billion trips. That is roughly ~38.2 million trips per day.2

Every single one of those trips required a localized auction, a routing calculation, a fraud check, and a payment settlement - all happening in under 500 milliseconds. This is not “hailing a cab.” This is high-frequency trading applied to the movement of atoms.

> root/silicon/geometry/h3_index

2. The Geometry of Efficiency (The H3 Index)

To understand the moat, you have to look at the map. But not the map you see on your screen. When Uber’s algorithms look at the world, they see hexagons.

In 2018, Uber open-sourced H3, a hexagonal hierarchical spatial index.34 This isn’t just a design choice; it is a fundamental engineering decision derived from topology.

Why Hexagons? If you try to analyze a city using a square grid (like a chessboard), you run into the “diagonal neighbor” issue.

• The distance to a neighbor sharing an edge is 1.

• The distance to a neighbor sharing a corner is √2 (~1.4).

This asymmetry makes the math for “smoothing” demand signals computationally expensive.5 A hexagon, however, is perfect. It has exactly six neighbors, and the distance to the center of each is identical.

This allows Uber to implement efficient convolution operations, essentially blurring demand signals across neighborhoods to prevent sharp “pricing cliffs”.

Uber “buckets” virtually all marketplace data (4.5 trillion records in their Gairos platform) into these H3 indices. This is data gravity. Any competitor trying to replicate Uber’s efficiency has to rebuild this density of data, hex by hex, city by city.

> root/silicon/infrastructure/gairos

3. The Nervous System: Gairos

To manipulate this hexagonal world in real-time, Uber needed a data platform that didn’t exist. So they built Gairos.6

Standard GIS tools are “stateless” meaning they see the world as a static snapshot. Uber needed a system that understands stateful stream processing.

Gairos is the “central nervous system” connecting raw data streams to decision engines. It handles the ingestion, indexing, and querying of spatio-temporal data at a scale that standard GIS tools cannot support.

Let’s revisit that “ride hailing query” or surge pricing phenomenon. When you open the Uber app, you aren’t just pinging a server. You are triggering a massive query against Gairos - their built sessions of stateful awareness (ie 6 cars in this hex, 2 empty, 3 dropping off within 60 sec and 1 en route)

“Return the count of open sessions and available drivers in H3 cells X, Y, Z (and their neighbors) over the last 5 minutes.”

Gairos executes this aggregation across sharded Elasticsearch clusters and returns the result in milliseconds:

• Data volume: It manages over 4.5 trillion records.7

• Throughput: It ingests over 1 million events per second.8

Uber's "surge" pricing is effectively an order book (bids/asks) that updates in real-time based on liquidity, identical to HFT algorithms, but applied to physical assets (cars/atoms) rather than stocks.

This infrastructure allows Uber to “statefully” track the city. They know not just where the cars are, but the velocity of demand relative to supply at a resolution of roughly 0.9 square meters.9


> root/silicon/algorithm/batching

4. The Marketplace Engine: From Greedy to Global

The second layer of the moat is the matching logic. In the early days, Uber used “Greedy algorithms.” If you requested a ride, the system instantly found the nearest driver. Simple. Fast. And mathematically wrong.

Greedy algorithms lead to local optimization but global inefficiency. They create “wild goose chases” where drivers cross paths inefficiently.

The batch matching paradigm is basically what modern Uber is using. Have you ever stared at the Uber app, watching the “connecting...” or “searching for drivers …” spinner for 5 seconds? It’s not slow. It’s thinking.10

In those few seconds, Uber accumulates a “window” of requests. It constructs a bipartite graph where one side is riders and the other is drivers. It then runs a massive optimization calculation (often min-cost max-flow) to maximize the global efficiency of the network.

It might make you wait an extra minute so that a rider 3 miles away doesn’t get stranded. This “preventative dispatch” maximizes total network throughput. Keep in mind that this feat is mathematically impossible for a smaller network.


> root/silicon/ai/deep_eta

5. The Brain: DeepETA & Transformers

To make this matching work, you need to know exactly when a car will arrive. Uber built DeepETA, a deep learning architecture based on transformers.11

Standard routing engines calculate ETA based on physics:
Distance / Speed = Time.
Uber’s DeepETA predicts the residual (the error) of that calculation based on context.

It uses vector encodings and “Self-Attention” mechanisms to understand high-order interactions. It knows that:

• “Heavy rain” + “Friday” + “London” = Gridlock.

• “Heavy rain” + “Sunday” + “Mumbai” = Minor delay.

Why Precision is Pricing?

This isn’t just about showing a timer on your screen; it is the fundamental anchor of the unit economics. Time is Money. If Uber underestimates a trip by 4 minutes, the “upfront price” quoted to the user is too low, and Uber pays the driver the difference out of pocket. If they overestimate, the price is too high, and the user churns to a competitor.

This system serves predictions with sub-millisecond inference latency.12 It is the heartbeat of the marketplace. Think about it: A 1% improvement in ETA accuracy saves tens of millions of dollars in “bad matches” and underpriced rides annually.


> root/silicon/ml_ops/michelangelo

6. The Factory: Michelangelo & Shadow Mode

Many companies claim to use AI. Uber has industrialized it. Running a model in a lab is easy. Running 10 million predictions per second while people are physically moving in cars is an engineering nightmare.

To solve this, Uber built Michelangelo, an internal Machine-Learning-as-a-Service (MLaaS) platform.13

The Feature Store (Palette)

To prevent “training-serving skew” (where the model in the lab works differently than the app), Uber uses a centralized feature store. It manages over 20,000 FEATURES, ensuring that the definition of “traffic density” is identical whether you are training a model or hailing a ride.14

Shadow Testing

How do you update the pricing algorithm without risking a total collapse of the marketplace? You use shadow mode. Before a new model goes live, it runs in the background for weeks. It receives live data and makes predictions, but its output is ignored. Uber compares the “shadow” predictions against the real-world outcomes to validate accuracy. Only when the shadow model outperforms the incumbent is the switch flipped.


> root/silicon/security/mastermind

7. The Immune System: Mastermind

Uber is a “trustless network.” You have millions of strangers getting into cars with other strangers, exchanging money instantly. If Uber relied on human review, they would be bankrupt. Instead, they built a three-layered digital immune system.

The Memory: Risk Entity Watch

Fraud is rarely a single event; it is a pattern. Uber built Risk Entity Watch to solve the problem of “amnesia”.15 Most systems check if a transaction is valid. Uber checks if the entity is anomalous. Using “entity feature generation,” it tracks the state of every user, device, and payment profile in real-time. It calculates signals like “ratio of chargebacks to rides over the last 90 days” on the fly. This prevents “account cycling”, even if a fraudster creates a new account, their device ID carries the “state” of their past crimes.

The Reflex: Mastermind

These signals feed Mastermind a high-speed rules engine.16 If a user tries to book a ride in London using a credit card issued in Mumbai that was just used in New York 10 minutes ago, Mastermind executes a “block” logic. It runs in <100 milliseconds, killing the transaction before the “request ride” animation even finishes.

The Investigator: Project RADAR

But what about new types of fraud? Enter Project RADAR.17

This is a “human-in-the-loop” AI system. It detects subtle anomalies that don’t violate specific rules yet (ie, a sudden spike in 5-minute rides in a specific suburb at 2 AM). It queues these suspicious clusters for human analysts. Once an analyst confirms the new fraud pattern, RADAR auto-generates a new rule for Mastermind.

This creates a compound radius of trust. Every attempted attack effectively trains the network, hardening the moat for every legitimate user.


> root/capital/financials/ad_subsidy

8. The Capital Moat: From Atoms to Pixels

The technical moat feeds the financial moat. The narrative of “Uber burns cash” is dead. In Q3 2025, Uber generated 2.2 billion in FCF. And the runway is huge. So far this business segment has been largely overlooked.18

But the real story is the shift in quality of revenue.

• For a decade, Uber sold atoms (rides). This is a low-margin business.

• Now, Uber is selling pixels (ads). This is a software-margin business.

In 2025, Uber’s advertising run-rate crossed $1.5 billion.19

Because Uber knows exactly where you are going (high-intent data), they can show you a relevant ad (e.g., a tequila brand while you ride to a bar). Because DeepETA knows the exact second you arrive, the ad impression is perfectly timed, commanding a higher CPM. This improves the unit economics of that trip without charging you a cent more. They are using digital margins to subsidize physical logistics.


> root/silicon/future/robotaxi

9. The Robotaxi Paradox

This brings us to the elephant in the room: The Death by Robotaxi. The bear case is simple: Waymo or Tesla will launch a robotaxi, undercut Uber on price, and Uber goes to zero.

This ignores the physics of capital.

AVs have high fixed costs (hardware). In this model, utilization is king (aka the utilization trap).

• If Waymo buys enough cars to serve the Friday night peak, 50% of their fleet sits idle on Tuesday morning. They go bankrupt from depreciation.

• If they only buy enough cars for Tuesday, they fail to serve the Friday peak. Users leave.

The hybrid network solution is what I believe will become reality in the coming years as the AV market is quickly ramping up, but hasn’t reached the critical volume yet. It is the best defense mechanism that Uber has to be ousted by other players. So far, Uber has emerged as the only player capable of dispatching a Waymo when conditions are perfect and a human driver when it’s raining or demand is crazy.

• Base load: Robots handle the consistent miles.

• Peak load: Humans (who bring their own capital) handle the spikes.

The proof is in the contracts. In late 2024 and 2025, Waymo expanded its Uber partnership to Austin and Atlanta20, Cruise relaunched its autonomous fleet exclusively on the Uber platform21 and Avride deployed robotaxis in Dallas22. The hardware makers are realizing that it is cheaper to pay Uber a fee for instant access to liquidity than to spend billions fighting for users.


> root/signal/risk_assessment

10. A Note of Caution (The Anti-Thesis)

While the “Uber getting disrupted” scenarios are less probable today, we must remain intellectually honest. There are risks:

• The “Super App” threat: If Google integrates Waymo directly into Google Maps with a “Book Now” button, they solve the distribution problem instantly.

• Regulatory fracture: If cities ban mixed traffic or mandate exclusive AV zones, Uber’s hybrid advantage evaporates.

There is nuance here that deserves its own article. I want to explore these specific bear cases in a future SILICON post.

The Signal in The SILICON

I started this audit looking for a taxi company facing obsolescence. I found a logistics utility facing a monopoly.

Uber has digitized the chaotic movement of the real world. They have built a computational moat (H3, Gairos, DeepETA) and are effectively building what Packy McCormick (Not Boring by Packy McCormick),calls a "Logistics Protocol" for the city23 - a programmable layer for the movement of people and goods that third parties (like Waymo) must plug into.

and a financial moat (FCF, ads, equity stakes) that are incredibly difficult to replicate.

The “slow” app is actually a massive supercomputer solving a global logistics equation. Uber is transitioning from a low-margin operator of human labor to a high-margin dispatch layer for the autonomous world.

In a world of noise. Long the signal

Thanks for reading!

Please consider helping me out, by subscribing - It’s free!
Or share / restack the article!

Share

The Rabbit Hole Goes Deeper

I originally planned this as a single post, but the research forced my hand. The “Uber vs. Tesla” debate is too nuanced for a hot take. It requires a model.

I am dissecting the collision between two distinct economic realities:

• The aggregator (Uber): Network effects and liquidity.

• The vertically integrated (Tesla/AVs): Physics and unit economics.

This isn’t just about stock prices; it’s about platform architecture vs. hardware integration. I am prepared to turn this into a 3-part deep dive analyzing the math of disruption.

Do you want the full breakdown? Let me know in the comments.

Leave a comment

Operational Update: Calibrating for Signal

I am adjusting the parameters for the SILICON series.

The research required to properly audit these architectures (tech, physics, and financial reality) has evolved. Each deep dive now demands 35+ hours of research + synthesis + editing to ensure it cuts through the noise.

I am refusing to dilute the output for the sake of the algorithm.

The new cadence: 1-2 high-fidelity dispatches per month.
The goal: Zero fluff. Maximum density.

References

23

Vertical Integrators by Packy McCormick (Not Boring):

Not Boring by Packy McCormick

Vertical Integrators

Welcome to the 669 newly Not Boring people who have joined us since our last essay! If you haven’t subscribed, join 231,376 smart, curious folks by subscribing here…

Read more

a year ago · 226 likes · 24 comments · Packy McCormick

[Accessed at: 01/21/2026]