AI came to Davos. No one actually talked about AI

Davos has always loved a future you can point at. The lanyards multiply, the black cars idle, and the world’s most consequential people practice the art of sounding calm while describing the future as an inevitability they just happen to profit from. And this year, AI showed up at Davos with a price tag you can meter, a permit you can revoke, and a CFO who wants receipts.

The week’s scenery felt on-brand: a promenade plastered with tech logos, meeting rooms stuffed with executives talking about “transformation,” and enough earnest panels to drive a small hydropower station. But the AI conversation in the Swiss snow globe didn’t float around model benchmarks or sci-fi inevitability. It didn’t really float at all — landing, again and again, on constraints. Alongside power, land, data access, security, and governance was the question that follows every trillion-dollar buildout like a shadow in sensible shoes: Who, exactly, gets to scale?

At the World Economic Forum, AI showed up looking like a supply chain. Davos treated AI as a scarcity stack and dragged the whole conversation back down to Earth. 

And the result was an AI week with a new hierarchy. At the bottom sits the meter: electricity, grid capacity, cooling, and the data centers that turn capital into compute. In the middle sit the keys: orchestration layers, permissions, identity, data rights, compliance, and the corporate systems that decide which agents can touch which workflows. At the top sits the proof: ROI, measurable outcomes, and the kill switch that turns this from a moonshot into a budget line that survives another quarter. This week was all about the points where ambition hits friction and where the “platform war” starts looking like who has control of the bottlenecks.

And once you see it that way, you can also see why Davos felt more zero-sum than last year. Not because CEOs stopped believing, but because they started itemizing what belief costs.

The meter: power

You could walk the Davos promenade and get the impression that AI is a software story, because the signs always look like software. Then, you’d talk to executives for five minutes and realize Davos was mentally stuck inside a utility bill.

The most Davos thing Elon Musk did in his first Davos appearance was go on a solar tangent while talking about AI. He argued that “a small corner of Utah, Nevada, or New Mexico” could generate all the electricity the U.S. uses, then complained that “tariff barriers for solar are extremely high,” inflating the cost of deploying it. The point was less the geography lesson than the instinct behind it. In a room full of people who love frictionless narratives, Musk mentioned the friction — and circled this key point from the week: When the AI buildout gets discussed seriously, it turns into an argument about energy production and policy.

You heard that theme everywhere, even from leaders who prefer their constraints quietly handled by someone else’s department. Amazon $AMZN CEO Andy Jassy put it plainly: “There is a power shortage.” He talked about the AI labs “consuming gobs and gobs and gobs of power,” and he framed Amazon’s response as pure industrial scrambling: “We’re doing everything we possibly can to add power, going and investing in small modular nuclear reactors.” He wasn’t delivering a climate sermon or a visionary monologue. He was describing a binding limit: A global buildout hits the grid, hits permitting, hits supply chains, hits the clock. In a world where AI capability increasingly maps to energy access, the winners don’t just ship features; they secure electrons.

Microsoft $MSFT CEO Satya Nadella delivered the same theme with a more Davos-friendly phrase: The thing Big Tech is really burning is legitimacy. “We will quickly lose even the social permission to take something like energy, which is a scarce resource, and use it to generate these tokens,” Nadella said, unless the output improves real outcomes — “health outcomes, education outcomes, public sector efficiency, private sector competitiveness.” For the past two years, the pitch has been that society should reorganize around AI because it’s inevitable. Nadella is now saying AI must earn its place on the grid.

Jensen Huang used Davos’ favorite rhetorical move — turning a technology into an economy — and did it with hard-hat language. The Nvidia $NVDA CEO called AI “the largest infrastructure buildout in human history” and described “AI factories” as foundational, the kind of phrase that makes policymakers lean forward because it sounds like jobs, taxes, and national capacity. Infrastructure gets funded. Infrastructure gets defended. Infrastructure gets prioritized.

Nobody wants to sound like they’re asking for a bailout. Everyone wants to sound like they’re building a national asset. That framing invites a particular kind of scrutiny: the kind reserved for data centers, transmission lines, permitting fights, and price shocks. It also pushes the platform war down the stack. If electricity is the limiting factor, control starts with who gets to build and who gets to plug in. 

The keys: permissions, orchestration, and the enterprise

Once power is scarce, access becomes power’s best friend. Once AI is physical, it becomes governed. Davos was full of executives competing to define “the layer” that matters most — because the company that owns the interface and the permissions also gets to decide who thrives in the ecosystem, a fight about orchestration, permissions, and the systems that let agents touch real work without turning the company into a liability.

Workday $WDAY CEO Carl Eschenbach claimed his company is positioned to become “the front door to work.” Workday already sits on HR and financial data, already manages onboarding and access, already handles permissions and performance management for human employees. That scaffolding is the real product: AI agents add a new class of worker. The company that controls "the front door" controls entry, access, and accountability.

At Davos, there was a race among enterprise vendors to become that orchestration layer for agents. Salesforce $CRM described “forward-deployed engineers” embedded with 120 of its largest customers so it can learn the hard, bespoke lessons and ship them back as scalable products. Microsoft pointed to its own orchestration and data-access tooling as companies try to unify systems without migrating every dataset into one monolith. Snowflake $SNOW’s CEO admitted his biggest fear is speed — whether the company can move fast enough “before OpenAI or Anthropic move down the stack,” into data storage and access, and displace incumbents.

The key is that “agentic AI” becomes a governance question the second it touches something that matters. Paychecks. Orders. Compliance workflows. Customer records. Permissions turn into power. Orchestration turns into a moat. Meta $META chief technology officer Andrew Bosworth described the investment moment as “this tremendous land grab of power, data centers, and GPU capacity.” Land grabs aren’t just physical; they involve rules, access, and control.

The “keys” theme also showed up in a place Davos likes to pretend is separate from business. Geopolitics is permissions by other means.

Anthropic CEO Dario Amodei criticized the Trump administration’s deal to allow Nvidia’s H200 chips into China as “crazy” and “a bit like selling nuclear weapons to North Korea” — because the AI stack has a passport now. Advanced compute is being treated as strategic material, and the gatekeepers include governments. That redraws the platform war. The question is no longer only who builds the best model. The question becomes who can guarantee access to the inputs that allow models to run at scale, who can operate inside the compliance perimeter, and who can survive the next tightening of rules. A company can lose on intelligence and still win on permissioning. Davos understands permissioning.

Jim Hagemann Snabe, Siemens’ chairman, took the governance point and pushed it into executive behavior: CEOs need to be “dictators,” he said, about where AI gets deployed and how initiatives get forced through the company. Tools don’t transform organizations. Power does. AI at scale requires decisions that create winners and losers inside companies — and leaders would prefer that to happen quietly, without a thousand local vetoes.

The proof: ROI

Once the meter constrains and the keys govern, proof becomes the filter. Julie Teigland, EY’s global vice chair, put the ROI logic directly on job design: “There is no ROI if you’re not willing to change the job descriptions.” She’s talking about training, role redesign, and the hard organizational work that most “AI transformation” decks politely skip. She even warned that getting stuck in endless pilots becomes a “death trap.” That’s Davos language for your board is going to lose patience. Teigland described the shift as “more real” as AI moved “from hype to scale.” The “scale” phase comes with finance-committee oversight, procurement rules, security reviews, and the oldest corporate question of them all: show me.

This part of the story is where ROI becomes the enforcement mechanism. ROI is the thing that decides which vendors keep getting budget, which transformations keep getting headcount, and which AI initiatives get filed under “interesting learning.” The “keys” fight has the same feel as every prior platform fight — except now it’s happening inside regulated, risk-managed institutions that don’t love surprises.

And Davos people don’t just fear surprises. They fear audit trails.

Even the AGI discourse — the part of the conversation that usually floats above practical constraints — carried a countdown clock. Demis Hassabis said AGI is “still 5–10 years away,” the kind of timeline that encourages investment, because it says the prize is real but also encourages discipline, because it implies a long middle stretch where companies must prove value before the future arrives. When the buildout is expensive, the grid is finite, and public patience is conditional, the companies that scale are the ones that can show outcomes.

The bubble question, politely kept off the main stage

One of the more revealing Davos tells was how rarely people wanted to say “bubble” out loud. Nobody wants to be the person suggesting the party is fragile while they’re still trying to get invited to the next dinner.

BlackRock $BLK CEO (and WEF interim co-chair) Larry Fink said during an interview with Bloomberg TV that he “sincerely believe[s] there is no bubble in the AI space” — although he does expect “some big failures” and “huge winners... and losers.” But the size of the buildout, he said, will drive huge AI growth. That’s great to hear if you’re on Wall Street, but the anxiety leaked through anyway — mostly through historical analogies and moral-permission arguments.

Bosworth, Meta’s chief technology officer, reached for the comforting infrastructure comparison: “We don’t regret railroads, telecom fiber … all the build-ups of this kind that we’ve done in history, we have ended up feeling great about.” He then described the present moment as “this tremendous land grab of power, data centers, and GPU capacity.” That’s a defense, yes, but it’s also an admission: The fight is physical, and the scale is absurd.

Davos, in other words, behaved like Davos. It kept the existential dread in the background and talked about execution in the foreground. That’s what happens when the people speaking are the ones signing checks — and when the checks are funding concrete, transformers, and nuclear-reactor conversations that don’t fit neatly inside a keynote. AI showed up this week as a scarcity stack — and scarcity changes behavior. It makes everyone care about the little things.

The platform war sounded less like a beauty contest and more like a contest for choke points. In this version of the story, the winners won’t just have the best model. They’ll control the meter, hold the keys, and produce the receipts. AI didn’t get quieter in the Alps. It got governed. It got physical. It got competitive in the specific way markets get competitive when the inputs become scarce, and the scoreboard becomes cash flow. AI’s future still showed up with all its old confidence. It just also showed up with a utility bill, a compliance checklist, and a demand for receipts.