The AI Reckoning
When the supply side admits the demand side was right
Satya Nadella just said the quiet part out loud. At Davos this week, Microsoft's CEO warned that AI risks losing "social permission" to consume energy if it doesn't deliver real benefits. This from the man whose company is spending $80 billion on AI data centres. The cognitive dissonance is staggering—until you realise he's been watching his own product fail internally.
The Copilot That Couldn't
Three weeks before Davos, Nadella emailed engineering managers about Microsoft's flagship AI product. He told them Copilot's integrations with Gmail and Outlook "don't really work" and are "not smart." This wasn't a pep talk. Nadella has essentially become Microsoft's chief product manager, holding weekly sessions with engineers and personally recruiting talent from OpenAI and DeepMind.
Meanwhile, Microsoft has been adding Copilot to Notepad—a text editor that's remained largely unchanged for decades. LG had to announce it would let customers delete the Copilot shortcut from their smart TVs after backlash.
The core problem: Companies are slapping chatbots on top of existing products rather than doing the hard work of deep integration that actually delivers value.
The Numbers Don't Lie
PwC's 29th Global CEO Survey, released at Davos, surveyed 4,454 CEOs across 95 countries. The findings are damning:
CEOs reporting both cost and revenue gains are two to three times more likely to have embedded AI extensively across products and services—not just deployed a chatbot. The pattern is clear: off-the-shelf AI solutions bolted onto existing workflows don't work. Deep integration does.
The Klarna Effect
No company illustrates this better than Klarna. In 2024, CEO Sebastian Siemiatkowski declared that AI can do all human jobs and slashed 700 customer service positions. The company partnered with OpenAI, routed three-quarters of customer interactions through bots, and celebrated $10 million in savings.
By mid-2025, Klarna was rehiring humans. Customer complaints had surged. Satisfaction ratings dropped. The AI couldn't handle nuance, empathy, or complex problem-solving.
Siemiatkowski's admission: "Cost unfortunately seems to have been a too predominant evaluation factor... what you end up having is lower quality."
The Burn Rate Problem
While companies struggle to extract value, AI providers are haemorrhaging cash at unprecedented rates.
Even OpenAI's head of Sora admitted the economics are "currently completely unsustainable." CEO Sam Altman acknowledged there's no ad model that can support people making funny memes to send to three friends.
The bailout attempt: OpenAI's CFO suggested seeking a federal "backstop" for chip investments. David Sacks, Trump's AI czar, responded: "there will be no federal bailout for AI." The comments were walked back within hours.
Economist Sebastian Mallaby projects OpenAI could run out of cash by mid-2027. The company has committed to $1.4 trillion in data centre infrastructure. Mallaby's assessment: an OpenAI failure wouldn't be an indictment of AI—it would merely be the end of the most hype-driven builder of it.
The Efficiency Revolution Happening Elsewhere
While US companies burn billions chasing marginal improvements, Chinese labs are proving you can build competitive models for a fraction of the cost.
SemiAnalysis estimates DeepSeek's total infrastructure investment exceeds $1.3 billion when accounting for GPUs, R&D, and operations. But their inference pricing tells the real story: what costs tens of thousands per month on closed APIs can be done for hundreds of dollars on DeepSeek
The architectural innovations driving this will be adopted by Western labs. But they'll be playing catch-up while still justifying massive infrastructure spending that may have been unnecessary.
What the 12% Know
The PwC survey's most useful finding isn't that 56% see no benefit. It's that the 12% seeing both cost and revenue gains share specific characteristics:
Embedded AI extensively across products and services
Established responsible AI frameworks
Built technology environments enabling enterprise-wide integration
This is the same pattern we've seen with every technology adoption cycle. The companies extracting value aren't buying enterprise software subscriptions and telling employees to "go forth and use it." They're redesigning workflows, retraining teams, and treating AI as a transformation project rather than a procurement exercise.
PwC Global Chairman Mohamed Kande: "Somehow AI moves so fast that people forgot that the adoption of technology, you have to go to the basics."
The Reckoning Ahead
We're entering what Kande calls "the most testing moment for leaders." CEOs must simultaneously run their current business, transform it in real time, and build entirely new business models. Most don't have a clear plan.
Nadella's "social permission" warning deserves attention. Society is being asked to accept enormous energy consumption, job displacement anxiety, and an avalanche of AI-generated content. If the benefits don't materialise beyond a handful of tech companies, the backlash will be severe.
The irony: AI genuinely is useful. We use it daily for tasks that would have taken hours. But the gap between what AI makes possible and how companies are deploying it remains vast.
The technology works. The implementations mostly don't. Deep integration wins.