For me, the last two weeks were all about the future of telco. The rest of the world had its own headlines, some of them bigger and a good deal stranger. Mine were an incredibly deep dive into transformation, AI, and where operators fit in what comes next. First the fourth edition of CASA, our CPaaSAA event in Amsterdam, where GSMA was a strategic partner — an event that takes an enormous amount of energy to create and to run. Two days later I was on a plane to Silicon Valley for the first Nova Future Summit in Napa, which GSMA runs together with the venture firm Wing. As I write this it is Thursday, 1 October. A week ago we were wrapping up CASA. Since then I have done Nova, spent time in Santa Clara, and today I went whale watching in Monterey. Two events, one week apart, with different formats, different locations and different strengths.

Four editions in, we are starting to understand what CASA is for. It was never only an operator event. It brings operators together with the cloud comms players — CPaaS, CCaaS, UCaaS — and increasingly with AI and tech companies, in a room small enough that people stay for the whole discussion and say what they actually think. My takeaway this year was trust: the case for operators as the trusted foundation underneath AI, and what it will take for them to stop talking like a supplier.

Nova did something that is harder to do in Amsterdam. Being in the Bay Area, it drew investors and AI startups in real numbers — five hundred people from telecom, technology and capital, under Chatham House Rule. My takeaway there was scale. AI is huge, it is coming at operators fast, and the operators know it. A detour to Santa Clara to spend time with NVIDIA’s telco team only underlined that.

What I missed in Napa was the layer in between. A couple of cloud comms players were there — mostly people I already knew — but the big names in UCaaS and CCaaS were absent, and cloud comms was not on the programme. Operators, hyperscalers, chipmakers and investors, yes. Not the companies that have spent twenty years building on top of operator networks and turning operator assets into products enterprises buy.

You could feel it. Most people understood their own half of the picture far better than the other. The telco people knew networks, regulation and compliance. The tech people knew models and infrastructure. How the two connect — where each piece actually sits — was the conversation nobody quite had.

The same gap runs through the best piece of AI thinking published this year. Konstantine Buhler’s The Cognitive Revolution, published by Sequoia on 2 September, opens with him riding to work in a Waymo and noticing that the trip asks nothing of his muscles and nothing of his mind. It resists the singularity talk and reaches for history rather than prophecy. It is very good.

It also has a hole in the middle. Buhler thinks about where intelligence lives at two scales — the nation and the pocket. He says almost nothing about everything in between. And everything in between is where the money is.

1. Thinking is about to get cheap — and demand will flood in

Buhler splits all work into physical and cognitive. Over two centuries, physical work went from almost entirely biological to almost entirely mechanical; cognitive work is on the same curve, only faster. And when the price of something collapses, demand does not shrink to fit. It floods. Jevons saw it with coal in 1865 — more efficient engines multiplied Britain’s coal consumption rather than cutting it. Buhler applies the same logic to thinking. Most problems on earth go un-thought-about today, not because they are unimportant but because thinking is expensive. The doctor who never gets to the latest literature. The small business with no budget for an analyst.

Make cognition nearly free and all of that arrives at once. So where does it run?

2. He has the nation and the pocket — nothing in between

To be fair to Buhler, geography is not absent. At the top end he has it: energy and compute are the new raw materials, intelligence is the finished good, and nations that once measured power in refining capacity will measure it in grid capacity. At the bottom end he has it too. Explaining why home and work are recombining, he writes that “the machine now lives everywhere” — one of the most powerful tools in history, sitting in your pocket.

Read that line again, because it is the assumption the whole industry is running on. Intelligence as placeless. Summoned from nowhere in particular, appearing wherever you happen to be. That is how software has behaved for twenty years, and we have stopped noticing that it is a design choice rather than a law of physics.

The Industrial Revolution did not work like that. Coal was somewhere. Rail went somewhere, and the towns it reached prospered while the towns it missed did not. The fortunes accrued to the people who owned the ground the material crossed as much as to the people who invented the machines. Buhler’s titans of this era — Nvidia, TSMC, the model labs, the hyperscalers — are Carnegie and Rockefeller, updated. But that is a list of who makes the material, not a map of where it moves.

The nation is the flag on the map. The pocket is the destination. The layer in between — networks, regional compute, the premises of regulated enterprises — is the route. And the route is where the toll gets collected.

If that sounds abstract, look at what Stripe just bought. OpenRouter does not build models and does not build applications. It decides which model gets called for which job. It launched in 2023, and in August Stripe agreed to acquire it. Note who the buyer is: not an AI lab, not a hyperscaler, but a payments company. Stripe built its business by sitting where money moves. Now it wants to sit where AI spend moves too. The models are commoditising. The router is what got bought — by a company that knows exactly what a toll booth is worth.

3. The barbell — and where it breaks

So why would one of the sharpest investors in the Valley skip the middle? Probably not by accident. Seen from Sand Hill Road, AI looks like a barbell: the model labs and hyperscalers at one end, the device makers at the other, trillion-dollar companies on both, and everything in between a pipe.

I run into this view constantly, and it is not a stupid one. The last twenty years support it — the value went over the top of the operators, not through them. The middle is capex-heavy, regulated and local, which is not where venture returns live. Sequoia is not in that business, so from where Buhler sits the barbell is a rational view. It also makes life simpler: two places to put your money, and nothing in between worth understanding.

But the barbell only holds if nothing forces value into the bar. Five things do.

Regulation. The barbell was built for consumers. A bank, a hospital or a ministry cannot send everything to a hyperscaler in another jurisdiction, and it cannot run its risk on a phone. GDPR, the EU AI Act, data residency and sector rules are not friction on the way to the barbell. They are the reason it bends.

Containment. Agents act. An agent that moves money, books care or speaks to a customer on a company’s behalf has to run somewhere it can be governed, logged, audited and, if necessary, switched off. That place is neither the model lab nor the handset. It is the layer in between.

Trust. Is this number really who it claims to be? Has this SIM just been swapped? Is this device where it says it is? Those signals come from the network. Apple and Google compete here, with passkeys and device checks, and that is fine — but a device check proves who is holding the handset, not that the number behind it has not been hijacked. And the handset is a consumer product. Enterprises do not buy their AI from the company that made their customer’s phone. AI that acts on people’s behalf will need the network’s answer on every interaction that matters.

Power. Buhler notes that intelligence per watt is getting cheaper fast. Jevons says that is not relief: cheaper per unit means vastly more units. Inference at scale is a power problem before it is a compute problem, and power is the most stubbornly local commodity there is. It has a grid, a connection queue, a regulator and a national politics. You cannot ship it to where the demand is. You put the demand where the power is. That gives the map a regional shape, just as coal did: inference will cluster where there is grid capacity to spare, and whoever already holds sites, power and connectivity in those regions has a head start.

Load. This one is closest to a law of nature. Agents do not behave like people. They do not make a call, browse for a few minutes and go quiet. They run continuously and in parallel, on millions of phones and devices at once, calling models, tools and other agents, and pushing signalling and traffic through the network in patterns it was never designed for. It came up very directly in Santa Clara: if agentic AI is going to run on every device, the network underneath has to be built to carry it. In barbell terms, the weights at both ends are getting heavier by the month. Sooner or later the bar has to be upgraded — or it bends.

The split between consumer and enterprise is the heart of it. As a consumer, I am happy to run my agents on my phone. It is massively powerful and I trust Apple to secure it for me. An enterprise is a different buyer. It wants compute where its data is created and held: in a box under the desk — NVIDIA now sells one built precisely to run agents there — in its own data centre, in a regional facility, or inside an operator’s network. That last option is not guaranteed, easy or trivial. But it is a viable play, and it sits in the bar, not at either end.

Geography cuts in more than one way, too. In the US, the hyperscalers and model labs are local and broadly trusted; the debate there is about national security rather than sovereignty. Outside the US it is a different conversation. Enterprises and governments are pushing back on running sensitive workloads in American model labs and hyperscaler clouds. Whether that demand lands with a European hyperscaler is far from certain. It may well land at the edge instead.

We have been here before. Mobile edge computing was supposed to bring compute into operator networks years ago, and it largely stalled, because the hyperscalers were better, faster and cheaper and nothing forced the issue. What has changed is everything that now pushes back: sovereignty, latency, data privacy, control, power. That is the pendulum I described at CASA26. For twenty years it swung toward the centre, and investors like Sequoia did remarkable things riding it. The next twenty will be structurally different, with compute moving to where the data sits.

So Buhler is right about the ends. The labs will be enormous, so will the device makers, and for consumers the barbell may well hold. But enterprise is a different play — different buyers, different rules, different risk. Nobody knows exactly where it ends up, and I do not claim to. But calling it a barbell is premature. The barbell is a picture of the last cycle, not a law of markets. Value goes to the ends when the middle is interchangeable, and back to the middle when the middle becomes the constraint. Regulation, containment, trust, power and load are turning it into exactly that.

Cognition has coordinates. Which means somebody owns the ground it runs on.

4. Three contenders for the route — and the layer that connects them

As inference moves out of the hyperscaler data centre, three kinds of player are competing to host it.

Telcos are slow, struggle with investment decisions and carry the most legacy. But they are distributed, they own the radio network, and they know regulation, privacy and compliance better than anyone else on this list.

Neoclouds move at a speed telcos can only envy. But regulation is new to them, and neoclouds come and go — a hard sell to a regulated enterprise.

Content distributors already have locations everywhere, built to push video for big brands, plus the software to match. Last week Akamai showed what that footprint is worth: an $11.6 billion, seven-year compute commitment from Anthropic, for capacity across its distributed cloud. The first big toll on the route did not go to an operator.

These three are less competitors than layers. Capacity can come from any of them — Deutsche Telekom’s AI factory in Munich proves operators can build it too. Capacity is not the scarce part. The scarce part is being the party an enterprise trusts to decide where a regulated workload runs, and wrapping that in something it will actually buy.

Which is why the most important player is missing from that list. The last time a route got monetised, telcos owned the ground and cloud communications players built the layer on top — the APIs, the contact centres, the collaboration tools enterprises actually bought. The operators supplied the network. The cloud comms players collected the toll. Without that layer in the conversation, the telco case for AI drifts toward networks, 6G, defence and security. All real. But the argument for operators as the foundation — trust, identity, sovereignty, compliance, the regulated conversation itself — needs the layer that knows how to sell it.

5. The application Sequoia named — and walked past

Listing candidates for the defining application of this revolution — the automobile rather than the spinning jenny — Buhler names science at machine speed, the personal agent that runs your life like a chief of staff, and something in how humans connect and coordinate. Then he moves on.

That third one is my industry, and has been my life for the past twenty years — named in passing and left unexplored. Not the whole world, but a substantial piece of it: telcos beyond the data pipe, contact centres, cloud communications — every place where people talk to people, and to businesses. Communications is where intelligence meets an actual human being with an actual problem, in real time, under rules about what may be recorded, stored and processed. It is regulated by definition, because it has been around for a very long time.

This year marks 150 years since Alexander Graham Bell made the first telephone call. The barbell describes twenty years; communications is a 150-year business. It has been switched, digitised and rebuilt more times than anyone can count, and it is still here, because people like to talk. As I think this through, I am sitting in a restaurant in Monterey, music playing, a glass of wine on the table — and everyone around me is doing exactly that. Talking to each other.

For a while it looked as if messaging would replace voice. It did not. Voice is back, and AI is the reason. People talk to these systems, and every conversation can now be searched, analysed and turned into intelligence rather than disappearing when the call ends. KPN and ElevenLabs showed a real product on stage in Amsterdam: AI built into the mobile service people already use, as a small monthly add-on. They are far from alone. Deutsche Telekom is rolling out an AI call assistant inside its network, Rakuten has built AI into its calling app, and operators everywhere are racing to put AI into voice for consumers and small businesses, and to put a price on it. Some of that is hype, and making real money from it will be hard. The technology is the easy part. The hard part is product and go-to-market.

That is what sits in the bar. Not a pipe with no value, but one of the most structural, most regulated and most human layers of the whole economy.

6. Why the middle lost last time — and how it wins this time

There is a reason the barbell looked right for twenty years, and it is not that the middle lacked assets. The middle sold capability while the ends sold outcomes. Operators sold minutes, gigabytes and speeds. The hyperscalers sold developers a faster way to ship. The device makers sold consumers a better life in their pocket. Value flows to whoever owns the job the customer is trying to get done, and for twenty years that was rarely the operator.

The same trap is open again. The instinct is to lead with the next capability — 6G, the next network API, the next standard. Designing the next generation is necessary engineering. But leading with it is the logic that turned operators into the pipe in the first place: improve the product, then go and find the customer. Monetise 5G before you sell anyone on 6G.

Network APIs are the test case. The third day of CASA26 belonged to GSMA’s Open Gateway team, and it mattered, because these APIs are part of the trust layer every AI deployment will need. Not all of them — the ones gaining traction are the ones that answer a trust question only an operator can answer. Andrew Collinson, who heads research at CPaaSAA, made the case on our stage as clearly as I have heard it: if the industry does not make that first horizon — security, identity, trust — a commercial success, it will never succeed at the second. Treat it like an MVP. Make the handful of APIs that work pay, iron out the kinks, and prove to the operators who have invested that there is a business worth investing more in.

Language follows the same logic. At CASA26 we argued for retiring “network APIs” in favour of what they deliver: network signals and trust, and network quality. Nobody wants quality on demand — you always want quality. Quality assured in the moments that count is something a buyer understands. A capability is something you have. An outcome is something somebody pays for.

This is not a marketing problem. It is an organisational one, and there are signs it is being taken seriously. I am starting to see operators put commercial leaders in charge of product. The chief strategy officer of a large European telco put it as plainly as I have heard anyone: change the organisation, build new products, market focus before tech focus. That is how the middle stops being a pipe — not by owning more infrastructure, but by owning more of the outcome.

7. The bridge already exists

Operators do not have to learn this from scratch. For twenty years cloud comms has built on operator networks and shown what can be created from operator assets: the products, the pricing, the release-early habit, the go-to-market. And the collaboration is already under way. Ericsson bought Vonage to put network capabilities in developers’ hands. In Amsterdam, a German operator presented its own CPaaS platform — why it built one, and how much less daunting that is now the technology is proven.

CPaaS players are bringing network signals to market alongside the hyperscalers’ developer platforms, and we see CPaaS as one of the most important routes to market those signals have. A new wave of startups is building the orchestration layer just above, deciding which signal, which model and which channel an interaction needs. And the same companies are turning their heritage into platforms for AI. Twilio is furthest along. Many of the companies on stage in Amsterdam are doing the same from very different starting points — from the contact centre and the business phone system to the specialists putting AI into voice inside the operator core.

That is the bridge, and it is my view rather than a consensus one. Operators bring the trust and the network. Cloud comms brings the products enterprises actually buy. AI brings the intelligence. Leave one out and you get either a network without a product or a product without a foundation. The two already work together; they should do it far more deliberately, and with AI on equal terms. If that partnership stays a buyer–supplier dynamic, operators will hold the best hand they have been dealt since the smartphone and still miss the toll — owning the land the railway crosses and settling for the rent on the right of way.

8. The second kind of unevenness

Buhler ends optimistically, and I think he is right to. He is honest that the transition will be unevenly distributed, and treats that as a labour question — who loses the job, how fast they retrain. There is a second unevenness, and it is about ownership. The people who did best out of the last revolution were rarely the ones with the best machines. They were the ones who understood the map: where the material was, where it had to go, and who controlled the ground in between.

Intelligence is about to become abundant. Nobody has decided where it lives. That decision is being made now, in rooms like Amsterdam and Napa.

We are already building on what both weeks taught us. Nova was genuinely insightful — the scale, the capital, the urgency — and it is helping shape where we take CPaaSAA next. It is good to see GSMA backing both. They are a partner for our events and a natural one, because they represent the telco industry: in barbell terms, the bar itself. We are already talking about what we do together next year, and there will be more of it.

When we started CPaaSAA in 2019, the idea was simple: put operators and cloud comms in the same room, because each needed what the other had. Seven years on, with AI in the room as well, it makes more sense than ever. Voice has carried human conversation for 150 years. It is not going anywhere. It is becoming the foundation the AI era runs on.

So, the barbell, head on. The view from Sand Hill Road puts the value at the ends and a pipe in between. But a barbell without a bar is two weights on the floor. The bar is what you grip, what carries the load, and what decides how much anyone can lift. That is where operators, cloud comms and the trust they hold together now sit. It is where this revolution gets decided.

The map is still blank. Let’s draw it together.