I Spent 30 Years in Telecom. AI Regulation Is Walking My Old Road.

I watched governments regulate a technology industry from the inside — not as a scholar, but from a management chair. Three decades in telecom, including as CIO of Freedom Mobile. If you want to understand how AI will actually be governed, don’t start with the AI Act. Start with spectrum.

The air was for sale

Spectrum — the radio frequencies your phone uses — belongs to the public. But you can’t touch it without a government license, and governments don’t give those away. They auction them. Canada’s spectrum auctions have raised billions; operators bid enormous sums for the right to exist in a market.

Before a telecom company lays a single cable or hires a single engineer, it has already paid the state for permission to compete. The license comes with conditions: coverage obligations, service quality floors, sometimes price constraints. The auction is the regulation. The capital outlay is the governance mechanism.

I sat in the rooms where those conditions shaped planning. Network investments, expansion priorities, where we built and where we didn’t — all of it bent around license terms negotiated years earlier in a process most consumers never see. The government’s power wasn’t in telling us what to do day to day. It was in setting the terms of market access before we started.

AI is about to learn this lesson.

Market access is the regulation

Anu Bradford gave this dynamic its definitive treatment in The Brussels Effect (Oxford University Press, 2020): the EU’s real regulatory superpower isn’t its bureaucracy, it’s 450 million consumers. Set the terms of access to that market, and global companies adopt your standards rather than run two systems. Bradford’s argument, grounded in competition law, privacy (GDPR), and product regulation, is that market size converts directly into rule-setting power — no enforcement army required.

The EU AI Act is the Brussels Effect applied to AI. It doesn’t try to out-build American labs or out-spend Chinese ones. It says: if you want to sell here, these are the terms. And because running a separate “EU-safe” AI stack is prohibitively expensive, the bet is that companies converge on one standard globally.

That bet might not pay off — AI models are more portable than physical products, and the Act’s enforcement is still being built. But the mechanism Bradford describes is real. I felt it from the operator’s side: when market access has a compliance price tag, compliance becomes product strategy.

Capital was always the co-regulator

The legal scholarship tends to miss this, and three decades in management made it impossible for me to miss: governments don’t govern technology alone. Capital co-governs.

In telecom the dynamic was out in the open. Spectrum auctions meant only the well-capitalized could play. License conditions steered investment — rural coverage obligations pushed capital into uneconomic areas; wholesale access rules decided whether new entrants lived or died. The regulator drew the frame. Capital markets decided what got built inside it.

AI rhymes. Venture funding decides which paradigms get explored — the concentration of frontier-model investment in a handful of labs isn’t an accident, it’s allocation. Public-market scrutiny already does governance work: once investors started pressing on AI risk in earnings calls, “responsible AI” moved from ethics page to board agenda. No law changed. The cost of capital moved.

That’s governance without a regulator. Fast, powerful — and its values are growth, not safety, not equity, not truth. Any AI governance design that ignores the capital layer is designed for a world that doesn’t exist.

Three experiments, plainly stated

The usual framing — EU regulates, US innovates, China directs — is too tidy. But the divergence is real:

Europe is testing whether market access converts to standard-setting. The mechanism is sound in theory. In practice, AI’s portability and the Act’s slow enforcement mean we won’t know for years if the Brussels Effect holds for foundation models.

The US governs through industrial policy and procurement. The CHIPS Act, compute infrastructure bets, federal purchasing power — less visible than legislation, arguably more consequential. It also swings with administrations, which makes it powerful and brittle at the same time.

China fuses state capital with national strategy. The most upfront about the link between funding and control — and the clearest proof that whoever pays for the frontier shapes its values.

My instinct from telecom: watch procurement and licensing, not press releases. In every market I worked in, the binding constraints were never the speeches. They were the license terms and the auction prices.

What should actually be regulated? There is no air.

Spectrum worked because the air is finite, rivalrous, and publicly owned. AI has no air. So what do you license?

Not the math. “AI” is too broad to license — like licensing electricity. What you can license are the chokepoints: the scarce, rivalrous inputs without which frontier AI doesn’t happen.

Compute is the closest thing to spectrum, and it’s already regulated like it. US export controls on advanced chips to China are spectrum-style governance run through trade law instead of auctions: permission required to play. The structural problem: no government owns the chip supply chain. Spectrum was domestic and public; compute is TSMC, Nvidia, and a global web of dependencies. The leverage is real but messier and less accountable than a clean licensing regime — gates buried in trade bureaucracy instead of written in law.

Energy is the lever nobody talks about. You can’t train a frontier model on a laptop. At scale, AI is an energy business that happens to produce intelligence, and grid interconnection queues are already the binding constraint on data center buildouts in many regions. Energy is rivalrous, local, and the state already holds the switch. A government that gates grid access for massive training runs governs AI deployment without writing a single “AI law.”

Climate tokens miss the target. Pricing carbon per training run handles the environmental externality, not the capability question. It taxes the electricity, not the intelligence. Reasonable climate policy; useless AI governance.

“Free for all who can get chips” is what we have, and it isn’t free. Export controls, capital concentration, and energy constraints already gate the frontier. The choice was never regulation versus freedom — it’s explicit, democratic gates versus opaque ones.

So: regulate the chokepoints, not the math. License advanced compute at scale and grid access for massive training — the rivalrous inputs. Handle everything else at the application layer, the way we govern every other general-purpose technology. The air isn’t for sale here. The power lines are.

What I’d tell a policymaker

I advise leaders on AI strategy. If a government asked me tomorrow, I’d say three things:

1. Anchor regulation in something companies can’t route around. Paper rules without market-access leverage are theater. Spectrum taught me: the license is the law that matters.

2. Get skin in the capital game. The labs with the most compute set the frontier, and whoever funds the frontier shapes its values. Public investment in AI infrastructure isn’t nostalgia — it’s governance.

3. Expect fragmentation, plan for it. Divergent regulatory models mean divergent AI ecosystems. Policy written for one jurisdiction is policy written for a fragment. Build for interoperability or accept the split.

I spent my career inside the machine that AI regulation is becoming. The names change — spectrum, licenses, auctions become compute, standards, market access. The physics doesn’t. Whoever controls the terms of entry writes the rules of the game. Everything else is commentary I can bill by the hour.