The Bitter Lesson for Telecoms

The Bitter Lesson Telecom Is About to Learn

The winning network will begin with the outcome, not the network

By Dominic Endicott
Knowledge Towns | August 2026

In 2019, the computer scientist Rich Sutton wrote a short essay called The Bitter Lesson. It was aimed at the artificial intelligence community, but the lesson is about to arrive in telecom.

Sutton looked back over 70 years of AI research and found a repeated pattern. Researchers would spend years encoding expert knowledge into a machine. The results often looked impressive. Then a more general system, using more computation and learning, would overtake it.

The experts were not foolish. Their knowledge often improved the system in the short term. The problem was that their carefully designed rules could not improve at the same rate as computation, search and learning.

Telecom should pay attention.

The industry has spent more than a century engineering the network. It has become remarkably good at it. Modern telecom networks coordinate millions of assets, serve billions of people and deliver a level of reliability that most technology companies would struggle to match.

But the intelligence in these networks still largely reflects decisions made in advance. Engineers select architectures. Operators forecast demand. Standards bodies define expected behavior. Products are constructed around the infrastructure available to deliver them.

This made sense when spectrum was scarce, computing was expensive and mistakes in capital planning took years to correct.

AI changes the assumption underneath that model. It makes it possible for networks to observe what is happening, test different responses and learn where resources create the greatest value.

The bitter lesson, however, goes further than network automation. Telecom must eventually stop beginning with the network.

Customers do not want networks

There is a reason that five generations of mobile technology have struggled to produce five generations of commercial models.

Telecom companies sell connections, data, devices, capacity and service levels. These products are valuable, but they are rarely what the customer is ultimately trying to buy.

A hospital buying edge computing is probably trying to identify patient deterioration earlier or help clinicians make better decisions.

A manufacturer installing private 5G may be trying to reduce downtime.

A university looking at sovereign AI capacity wants to protect its research and give students and faculty access to advanced tools without losing control of its data.

The network matters in each case. But it is only one part of the answer.

This difference has enormous economic consequences. The value of avoiding a day of factory downtime may be many times the value of the connectivity involved. The same is true of preventing a hospital admission or helping a university turn research into a new company.

Telecom monetizes the input. Someone else usually understands the outcome, shapes the experience and captures most of the value.

We have seen this before. Telecom companies built much of the infrastructure of the mobile internet. The platforms above the network learned more about what people wanted and made the most valuable decisions. The operators carried the traffic while others captured the economic surplus.

AI could repeat this pattern on a much larger scale. This time, the decisions will reach beyond what appears on a screen. They will determine where computation runs, how energy is used, which infrastructure is activated and how intelligent systems act in hospitals, factories, universities and communities.

The critical question is whether telecom can become part of the intelligence that organizes the real world.

Physical AI brings intelligence back to place

The first phase of generative AI has been dominated by very large data centers. It is easy to conclude that intelligence will become more centralized and less connected to place.

Physical AI leads in another direction.

A robot moving through a factory, a clinical system supporting a physician and an autonomous vehicle travelling through a town all depend on the conditions around them.

They need more than access to a distant model. They may require local sensing, mapping, identity, resilient connectivity, nearby computation and a clear account of who is responsible when the system gets something wrong.

These requirements will differ from one place to another.

A clinical AI system operating in Burlington cannot be separated from the hospital’s data, workflows, privacy requirements, staff and physical infrastructure.

A manufacturing system in Maine must work with the equipment, skills, suppliers, power constraints and network assets that exist there.

This is usually presented as an edge computing opportunity. I think that understates it.

The opportunity is to connect the knowledge and needs of a place with the full set of resources that could serve them. Those resources include networks and computing, but also institutions, people, buildings, energy and local authority.

At Knowledge Towns, we call this the Knowledge Fabric.

The Knowledge Fabric would begin with a real outcome in a real place. It would then identify and assemble the resources needed to produce it. As the system operated, it would learn what was working and what was not.

The network would be one thread in the Fabric. An important one, but still a thread.

That is different from the autonomous network strategies now emerging across telecom. Those strategies generally start with an intention for the network: improve reliability, reduce energy use, allocate capacity or assure a particular level of service.

The Knowledge Fabric starts with an intention for the place.

How do we improve a clinical pathway?

How do we make a group of smaller factories more productive?

How do we give a university community secure access to AI?

The system then works backwards to the infrastructure.

The network has to become legible

Telecom has a credible claim on this opportunity.

Operators already own secure facilities, fiber, spectrum, towers, rights of way, power connections and thousands of locations close to the people and institutions that will use AI.

Many former central offices look like natural homes for distributed computing. They have fiber, physical security and, in some cases, useful power and cooling.

Yet it would be a mistake to announce that all these facilities are now edge data centers and begin another large capital program.

Some locations may be valuable for clinical inference, industrial systems or public safety. Others may have no local demand that cannot be served more cheaply from a large data center elsewhere.

We do not yet know.

The first task is to make the asset base legible.

What facilities exist? How much power is really available? What can be activated quickly? What are the security, regulatory and physical constraints? What would it cost?

The same discipline has to be applied to demand.

Institutions often say they want local or sovereign AI, but stated interest is not the same as a budget or a workload. Telecom needs to discover which uses genuinely require locality and which ones merely sound as though they do.

This is where learning matters. Its purpose is not to confirm the current edge thesis. A useful system should also kill weak ideas before large amounts of capital are committed.

Telecom has sometimes announced the architecture first and searched for demand later.

The Knowledge Fabric reverses the order.

Learning will happen at different speeds

There is a temptation to describe all of this as a rapid OODA loop: observe, orient, decide and act.

That is directionally right, but the word “rapid” needs care.

A network can identify congestion in milliseconds. A factory can measure downtime each day. A hospital may need weeks to know whether an intervention changed readmissions. A university may need a semester to see whether learning improved. A town may need several years to understand whether an investment changed its economy.

A Knowledge Fabric therefore needs several learning loops operating at once.

Fast loops can adjust infrastructure. Slower loops can test whether the experience improved. The slowest loops should shape major capital decisions.

The aim is to reduce the gap between an intervention and reliable evidence about its effect.

Nor should the language of experimentation become an excuse for carelessness.

Hospitals, vehicles and public infrastructure are not social media applications. People can be hurt.

Experiments must be bounded. Institutions must retain authority. Systems need secure data, clear accountability and the ability to reverse a decision. Autonomy can expand as evidence builds.

Security and sovereignty are not optional extras

Universities, hospitals and governments care where their data is held, where inference occurs and which laws govern the system.

They also care about what happens if a provider changes its terms, a cloud service fails or another country attempts to assert control.

Telecom companies understand many of these concerns. They operate regulated infrastructure and have longstanding relationships with governments and institutions.

But sovereign AI cannot simply become an expensive hosting product.

A regional university or community hospital may not have enough demand to support dedicated infrastructure. Ten institutions together might.

The opportunity is to combine demand without requiring each participant to give up authority over its data and decisions.

That suggests a federated model.

Infrastructure owners can make selected resources available without surrendering control of the underlying assets. Institutions can participate without putting all their data into one central pool. Learning can take place across the system while the rules remain visible and enforceable.

This institutional coordination may become more important than any particular network technology.

Prove one thread

I do not know whether telecom companies will seize this opportunity.

In truth, most probably will not.

The existing business is large. Reliability cultures are rightly cautious. Assets and information are divided across business units. Vendors control important parts of the stack. Large capital programs are easier to organize than small experiments crossing institutional boundaries.

The alternative is to pick one place, one institution and one important outcome.

Establish the baseline. Map the actual assets and constraints. Build the smallest credible intervention. See what happens.

A university might begin by providing secure AI capacity to researchers and local companies.

A hospital district might test whether local inference can improve one clinical pathway.

A manufacturing community might combine demand across several smaller firms that cannot justify advanced infrastructure individually.

Most pilots will reveal problems. Some will show that local infrastructure is unnecessary. A few may expose combinations of institutional demand and telecom assets that nobody planned in advance.

That is the point.

Over time, successful threads can be connected. The result will not be a single centralized system controlling a town. It will be a growing capacity for the place to learn, coordinate and adapt.

Phone networks carried conversations.

Broadband networks carried information.

The next network will help intelligence act in the physical world.

Telecom can become part of the system making those decisions.

Or it can provide the infrastructure while somebody else learns how best to use it.

That is the bitter lesson telecom is about to learn.

Dominic Endicott
Knowledge Towns
dominic@knowledgetowns.com