The bottleneck of autonomous driving is by no means AI.

An anomalous phenomenon is plaguing the entire autonomous driving industry: the most brilliant "brain" has been built, but the "nerves" of the physical system cannot keep up.

In recent years, the accuracy of perception algorithms for autonomous driving has increased nearly a hundredfold. Large models can detect a suddenly intruding pedestrian, a drifting plastic bag, or a construction cone ahead in just a few milliseconds. The cost of computing power has also dropped dramatically, with a chip costing just tens of dollars capable of running the full set of visual perception models. Capital, talent, and technology have flooded into the artificial intelligence track like a tide, bringing about visibly exponential progress.

However, when these intelligent vehicles actually drive onto urban streets, highways, and port terminals, the biggest source of risk that keeps engineers awake at night is surprisingly not algorithmic misjudgment, but a millisecond-level interruption of wireless signals at the most critical moment.

Why? Because the radio links carrying control commands, location information, and remote monitoring data are disturbingly fragile in the physical world, a long-overlooked aspect and a fundamental architectural flaw in the industry.

"Connection Failure"

Is More Unbearable Than "Recognition Error"

What exactly is the status of wireless communication in autonomous driving? It should be noted that autonomous driving systems are typical cyber-physical systems, whose safety boundaries are not entirely determined by on-board computing power, but are largely constrained by the integrity and timeliness of information. When an autonomous vehicle is traveling at 120 km/h, it covers a full meter every 33 milliseconds, meaning any end-to-end communication interruption exceeding 30 milliseconds will put the vehicle in a state of "blind flight" in physical space.

Let's compare two typical failure modes.

First, perception failure. The front camera temporarily cannot clearly identify the exact position of an obstacle due to strong backlighting, causing the confidence level of target detection to drop from 0.98 to 0.65. At this point, the system still knows it "cannot see clearly", and can adopt conservative strategies: reducing speed, increasing the following distance, requesting driver takeover (for assisted driving), or switching to redundant sensors. This failure is "self-aware", with clear boundaries and relatively controllable consequences. Engineers call this feature "graceful degradation" — the system cannot operate at full capacity, but will not crash suddenly.

Second, communication interruption. Remote control commands, remote monitoring data streams, and fleet cooperative avoidance messages of the vehicle are continuously lost due to sudden interference in the wireless channel. The problem is that such interruptions are often sudden: the channel may be perfectly fine one millisecond, but the next millisecond, the signal-to-noise ratio may drop below the threshold instantly due to the obstruction of a heavy truck's metal body or the sudden influx of hundreds of user equipment in the same frequency band. The sender may have no idea that the receiver has fallen silent, and the receiver does not know when the sender will resume operation. This failure is "unaware", with vague boundaries and unpredictable consequences.

From a risk perspective, perception failure is a high-frequency low-impact risk, while communication interruption tends to be a low-frequency high-impact risk.

For low-frequency high-impact events, conventional statistical methods and safety redundancy designs often provide insufficient coverage.

An autonomous driving company can easily calculate the average error rate of perception algorithms in millions of kilometers of testing, but it is difficult to measure the true probability distribution of catastrophic interruptions of wireless links in a complex urban electromagnetic environment.

This is because the triggering conditions for the latter are extremely complex: weather, building reflections, the number of concurrent terminals, load fluctuations of nearby base stations, and even solar activity can all affect that invisible electromagnetic connection.

More critically, these two risks have completely different impacts on the commercial closed loop.

Capital markets and insurance companies can accept that an autonomous vehicle occasionally brakes to a stop or pulls over due to perception difficulties — a scenario that also frequently occurs in human driving, with predictable claim costs.

However, if a Robotaxi without any backup control link suddenly loses connection in a busy urban area due to wireless interference, even for just three seconds, it will trigger the collapse of actuarial models, a full-scale tightening of regulatory scrutiny, and the erosion of public trust.

A single such incident is enough to undo the decade-long narrative of safety that the industry has built.

This is the most troublesome "black swan" effect in tail risk pricing: the event has an extremely low probability, but once it occurs, the losses are immeasurable. In this sense, the resilience of wireless communication, more directly than the precision of artificial intelligence, determines whether high-level autonomous driving can enter a positive commercial cycle.

Spectrum Is Scarcer Than Computing Power

But why can artificial intelligence advance by leaps and bounds, while wireless communication has been stumbling along?

There are underlying economic structural reasons behind this, starting with their completely different supply elasticities.

The core input factors in the field of artificial intelligence — computing power, algorithms, and data — have shown astonishing supply expansion over the past decade. Chip manufacturing processes have continued to advance along the inertia of Moore's Law, the construction cycle of 10,000-card GPU clusters has been shortened from years to months, the prosperity of the open-source community has allowed advanced algorithms to spread at almost zero cost, and the cost of obtaining high-quality labeled data has continued to decrease as automatic labeling technology matures. When the supply curve of a production factor keeps shifting to the right, its marginal cost approaches zero, and the ceiling of industry development is continuously raised.

Wireless communication is different, whose underlying constraints come from a naturally scarce resource that cannot be created through technological progress — the electromagnetic spectrum.

Spectrum is not computing power.

You cannot etch more usable frequency bands through TSMC's advanced manufacturing processes, nor can you generate a clean bandwidth out of thin air using a more sophisticated neural network model.

The propagation of electromagnetic waves in the physical world is governed by Shannon's theorem, free-space path loss, multipath effects, and interference superposition — all hardware laws that do not change with human technological enthusiasm.

Especially in the low-frequency bands suitable for connected vehicles (generally below 6 GHz), the scarcity of available spectrum resources is more severe than that of land in the core areas of first-tier cities.

Land supply can be increased through land reclamation or by raising plot ratio, but the physical capacity of spectrum has a clear upper limit: at the same time, in the same space, and on the same frequency point, only one set of signals can be reliably received.

This rigidity of supply is in sharp conflict with the explosive growth of demand.

A Level 4 autonomous vehicle generates dozens of terabytes of sensor data every day, and the amount of information that needs to be uploaded to the cloud in real time or shared with surrounding vehicles far exceeds the design capacity of current 4G/5G networks.

Not to mention drone logistics, remote operation, fleet coordination, and dynamic updates of high-definition maps — all these applications are competing for the same invisible electromagnetic space. When tens of thousands of intelligent entities operate simultaneously in the same urban area, spectrum congestion will no longer be an occasional interference, but a normalized environmental constraint.

The problem is that spectrum allocation mechanisms are far behind technological demands.

In most countries, high-quality spectrum still follows a decades-old "command-and-control" model, where regulators designate fixed uses and auction the spectrum to a small number of operators for long-term exclusive possession.

This rigid system results in large amounts of spectrum being effectively idle in both time and space dimensions, while users with real demand cannot obtain access rights. Although spectrum sharing technologies and dynamic spectrum access have been discussed for many years, large-scale commercialization has been constrained by existing interest patterns, regulatory inertia, and insufficient technological maturity.

While the computing power market is already enjoying on-demand elastic supply brought by cloud computing, the spectrum market remains in a pre-modern stage of "planned allocation" and "one auction, lifetime possession".

This well explains why autonomous driving companies can purchase the world's most advanced GPU clusters to train large models, but cannot use the same funds to obtain an absolutely reliable wireless link in harsh electromagnetic environments.

For things with zero supply elasticity, the price mechanism often fails.

Even with money, you cannot buy more spectrum, because spectrum is not a commodity, but a licensed right to use resources.

How Spectrum Governance Failure Undermines Commercialization

If we only view spectrum scarcity as a technical bottleneck, that would not be so bad. In reality, what truly makes wireless communication the shortcoming of autonomous driving is the institutional dilemma systematically ignored by the industry — the spectrum exhibits typical "Tragedy of the Commons" characteristics in scenarios where multiple active intelligent entities coexist.

What is the Tragedy of the Commons? For example, in a public pasture, each herder adds one more cow to graze for their own benefit, eventually leading to overgrazing that degrades the pasture, leaving all herders' cows to starve. The spectrum environment faces exactly the same dilemma when autonomous driving is densely deployed. Every operator wants to maximize the use of their wireless resources, transmitting higher-definition images, sending denser status messages, and pursuing faster response latency.

However, without effective coordination and priority allocation mechanisms, the reliability of all links will sharply deteriorate due to mutual interference.

The operating environments of high-level autonomous driving and drones are precisely high-risk areas for this Tragedy of the Commons.

Take an urban intersection as an example: multiple autonomous taxis of different brands, freight unmanned vehicles, roadside intelligent units, pedestrian communication devices with handheld terminals, Wi-Fi hotspots of shops along the street, and other conventional mobile communication traffic carried by nearby base stations may all be active at the same time and place.

These devices belong to different operators, service providers, and manufacturers, with no unified resource scheduling between them, and in many cases, they even treat each other as "noise". When the total load of this electromagnetic space approaches the capacity limit, any newly added transmitter may become the last straw that crushes all links.

The Coase Theorem theoretically provides a solution: as long as spectrum property rights are clearly defined and transaction costs are sufficiently low, users can achieve optimal allocation through private negotiation or market transactions. The operator that needs a reliable link more at a specific moment will bid to purchase the right to use the spectrum at that moment, allowing other operators to downgrade non-critical services.

However, transaction costs in reality are astonishingly high. Not to mention how difficult it is to complete the entire market process of discovery, negotiation, pricing, and settlement on a millisecond timescale, even the basic definition of property rights is impossible to achieve.

Existing spectrum licenses define frequency bands, geographical coverage, and authorization periods, but do not reserve interfaces for fine-grained usage rights of "specific time, specific space, and specific priority".

Moreover, traffic with different safety levels is treated equally and thrown into the same "best-effort" physical channel.

The braking pulse of an autonomous vehicle performing an emergency avoidance maneuver and the background traffic of another vehicle uploading entertainment system cache files have the same scheduling weight under the current network architecture.

This is equivalent to letting the siren of an ambulance and the sound waves of a car audio system compete fairly in the air.

From the very beginning of communication protocol design, priority and quality of service guarantee mechanisms were not prepared for such extremely heterogeneous scenarios.

What does this spectrum governance failure mean for the industry?

The key point is that it directly increases the institutional cost of large-scale autonomous driving operations.

A company can optimize all the hardware and software of its own vehicles, but cannot control the behavior of other entities in the same electromagnetic environment. When interference becomes an uncontrollable externality, the prediction of return on investment will be mixed with unquantifiable noise.

As a result, large-scale deployment will be delayed, either waiting for greater certainty in the institutional environment, or waiting for competitors to step through the regulatory minefield first. This state of mutual hesitation is the structural reason why many unmanned driving projects suddenly lose momentum when moving from "technical verification" to "commercial operation".

Why a Single Link Can Determine the Valuation of the Entire System?

The sign of any mature business model is obtaining stable pricing from the capital market. The prerequisite for stable pricing is that risks can be identified, measured, and managed. From this perspective, re-examining autonomous driving reveals a highly ironic mismatch: the industry has spent tremendous efforts to quantify, reduce, and insure AI decision-making risks, but lacks basic risk pricing tools for a more fatal variable — catastrophic loss of control caused by communication interruptions.

Generally speaking, in a system's loss distribution, most of the losses are mild and predictable, which form the "head" and "middle" of the curve; but at the far right (or far left, depending on the sign direction) of the distribution, there are hidden "tail" events with extremely low probability but extremely high losses.

A healthy financial system must conduct stress tests, set aside capital, and purchase reinsurance for these tail events, because over a long timescale, tail events are almost certain to occur.

The tail risks in the autonomous driving field are precisely concentrated at the moment when wireless communication loses control.

An autonomous vehicle performing a lane change while relying on real-time monitoring by a remote safety officer loses its control link at this moment. Even if the local intelligent entity attempts to execute a fallback strategy, the momentum in the physical world will not allow any room for error. The probability distribution of this risk is extremely difficult to estimate, as it depends on the number and power of instantaneous interference sources, channel fading characteristics, instantaneous base station load, the reaction delay of safety officers, and the actual execution effect of fallback strategies — each variable exhibits fat-tail characteristics.

This leads to a typical financial modeling dilemma: if both the probability of risk and the amount of loss cannot be reliably quantified, then insurance companies setting premiums, reinsurance companies accepting cessions, or investment institutions calculating discounted cash flows of enterprises will all fall into severe uncertainty.

This uncertainty has direct economic consequences: a sharp increase in capital costs. The financing costs, insurance costs, and compliance costs of autonomous driving companies will be systematically overestimated due to the tail risk of this fragile link.

The accumulation of these hidden costs may have already exceeded the explicit benefits brought by AI technological progress, which most business plans selectively ignore.

Therefore, those who truly understand the logic of tail risk pricing will reach a conclusion completely opposite to the mainstream narrative: the most effective way to increase the commercialization probability of high-level autonomous driving at present is not to continue stacking parameters in AI models, but to make the reliability of wireless communication links break through several key "nines" — that is, to increase from 99.9% to 99.999% or even higher.

Because each additional "nine" reduces the probability of tail risk by an order of magnitude, the risk premium of the entire system will drop sharply.

The benefits of this drop will definitely exceed the marginal improvement brought by increasing the object detection accuracy from 98% to 98.5%.

This does not mean that AI progress is unimportant, but there is a cruel law of diminishing marginal returns. After the perception ability of AI exceeds the average level of human drivers, continuing to invest massive resources in pursuing the last few percentage points of improvement will see the safety gains per unit of input sharply shrink.

In contrast, investing the same resources in the construction of communication resilience can reduce tail risks by several orders of magnitude more per unit of input than AI investment. A sensible resource allocation decision should concentrate resources in the direction with the highest marginal output.

Final Remarks

Having said all this, I am not trying to pour cold water on the booming artificial intelligence revolution. On the contrary, because I firmly believe that autonomous driving and unmanned systems will reshape the way human society operates, I am more eager for the industry to see this long-overlooked shortcoming.

Intelligent entities in the physical world cannot ultimately escape the constraints of physical laws. A car traveling on the ground at life-critical speeds, and drones performing missions over densely populated areas, rely not only on trillions of matrix multiplications per second for safety, but also on a "nervous system" that can withstand attacks even in an electromagnetic storm.

The difficulty of building this nervous system has been severely underestimated. It involves the institutional restructuring of spectrum resources, the thorough upgrading of communication architectures, the design of coordination mechanisms for multi-stakeholder games, and the entire industry's re-understanding of tail risk pricing.

The progress of artificial intelligence is predictable. With the continuous growth of computing power, continuous evolution of algorithms, and continuous accumulation of data, perception and decision-making capabilities will continue to improve.

However, the constraints of wireless communication will not be automatically lifted: spectrum will not naturally increase, interference will not disappear on its own, the Tragedy of the Commons will not resolve itself, and tail risks will not be reduced out of thin air. These problems require equal wisdom, resources, and institutional innovation to solve.

When the next critical moment for autonomous driving commercialization arrives, regulators, insurance companies, and the public will look past the dazzling autonomous driving demonstrations and focus on the liability document that must be signed. At that time, the most crucial question may not be "How smart is your AI?", but "Can your connection still protect the lives of everyone on board under the worst circumstances?"

The person or enterprise that can answer this question well, I believe, is the one that truly holds the key to the unmanned future.

This article is from the WeChat public account "Dongzhen Strategy", authored by Dongzhen Strategy, and published with authorization from 36Kr.