Podcast Summary: Azeem Azhar’s Exponential View

Episode: What it will take for AI to scale (energy, compute, talent)
Host: Azeem Azhar
Date: December 10, 2025

Overview

In this solo episode, Azeem Azhar explores what it will take for artificial intelligence to truly scale over the next two years. He examines the multiple layers of “absorption” required for AI to fulfill its economic, technical, and social potential—including the readiness of companies, the strain on power infrastructure, the competitive dynamics in the industry, political and sovereignty dilemmas, and shifting societal attitudes. Through this lens, Azeem frames the challenges and opportunities ahead, while answering live audience questions on strategies for organizations, model competition, value accrual, sovereignty, and security risks.

Key Discussion Points & Insights

1. The Central Theme: Absorption of AI

• Economic Absorption: Are companies and people meaningfully using AI, or is exposure mostly limited to casual chatbot queries?
  • “It’s almost trivially easy to put a product in the hands of ... millions of people ... But ... we need to go beyond someone downloaded the app and played with it a little bit.” (03:15)
• Organizational Challenges:
  • API access is easy; real transformation is hard due to organizational inertia and the need for new processes and mindsets.
  • Change is expected to happen faster than historic analogues (e.g., electricity), thanks to greater adaptability and developed playbooks.
• Physical Constraints—Energy & Compute:
  • The single biggest bottleneck now is power: Data centers are easy to build but much harder to connect to sufficient electricity grids.
  • Hyperscalers are securing massive electricity deals, showing demand outstripping supply.
  • “There is a scramble for getting energy into data centers. ... The demand is really high ... the hyperscalers are somewhat price insensitive in order to be able to build capacity.” (09:20)
  • Example: AWS had to turn down a major Fortnite contract due to lack of capacity.
  • Companies like Google and Microsoft are exploring fusion energy deals to secure future supply.
  • Efficiency improvements are accelerating: From 50 tokens per watt-hour (2022) to 600 tokens per watt-hour (2025), thanks to hardware and algorithmic advances.
• Resource Tension—Training vs. Inference:
  • Providers face trade-offs: Invest in training new models or servicing current customers?
  • “Model companies will be battling between where do they put their resources: into training the next model or serving company customers for revenues.” (13:30)
  • Analyst estimate: By the late 2020s, 70–75% of compute cycles will be devoted to inference rather than training.
• Long-Term Structural Shift:
  • The ongoing shift of business activity to computational and digital processes is seen as fundamental as the adoption of electricity (1880s–1930s).
  • “This is a fundamental shift in the economy. As fundamental as going from 1880 when nobody was really using electricity ... to the 1930s...” (11:45)

2. Economic Returns & Industry Timelines

• Uncertainty About Results:
  • Large investments are being made amidst great uncertainty regarding payback.
  • IBM’s CEO: Current data centers may cost $80 million per megawatt—significantly higher than estimates even a year or two ago.
• Why Mid-2026 Matters:
  • Enterprise AI rollouts really began in early 2024, following key releases from Microsoft and Google in late 2023.
  • “At that point [mid-2026] we should start to see more and more companies talking about the results they’re getting.” (19:10)
  • The expectation is that within two years, pilot projects should deliver visible ROI—or companies may quietly shelve them.

3. Political & Geopolitical Challenges

• Sovereign AI & Strategic Dilemmas:
  • For middle powers (not US/China), true sovereignty over critical AI infrastructure is increasingly challenging.
  • Nations like the UK, Gulf states, Brazil, India are investing in massive AI data centers—but chips and serving depend on global supply chains.
  • “It’s only sovereignty by vibes ... It kind of feels like sovereignty ... Because ultimately ... if you are building on the China stack [or] US stack, one of these two countries can say ... we don’t want to support you anymore.” (30:52)
• Collaboration is Needed:
  • Likely to see “minilateral arrangements” or regional alliances to share resources around compute, power, data, and talent.

4. Societal Attitudes and Legitimacy Crunch

• Public Ambivalence:
  • Despite 2 billion users of AI tools, public sentiment is increasingly pessimistic—“70–75% of Americans are pessimistic about what AI might bring” (citing Edelman Trust Barometer).
  • “It’s almost like you’re forced to use them ... because you need to participate. It’s an uncomfortable place to be.” (27:38)
• Grassroots Resistance:
  • Growing opposition to data center buildout—environmentalists and rural landowners are joining forces to block or slow projects.
  • 142 US projects (worth $64 billion) stopped since 2023.
  • Resistance is crossing political lines: described as “fanatically bipartisan.”
  • “It does introduce the idea of a legitimacy crunch ... how AI companies need to talk about what it is they’re doing ... talking about replacing every job ... may not be the messaging that is going to be appealing to anyone.” (29:40)

Live Audience Q&A Highlights

Should you jump into AI now or wait for stability?

• Azeem’s Advice: Start now. Waiting for stability means you’ll have to start from cold and miss crucial organizational learning.
  • “You should start now. You should have actually started a year and a half ago. So if you haven’t, you might want to just drop off this call now and get going.” (35:22)

OpenAI vs. Gemini—What’s next?

• OpenAI expanded quickly into many markets (sovereign deals, localization, enterprise) without always having clear product-market fit.
• Growth brings complexity and risk of losing focus, unlike more specialized competitors (Anthropic, ElevenLabs).
  • “You got product take-up before you’d understood why you had product-market fit.” (38:42)

Why didn’t the market react to OpenAI’s “Code Red” announcement?

• Markets are opaque, often acting on incomplete information; sustained trends matter more than short-term headlines.
  • “I simply wouldn’t ... count that team out and I wouldn’t count a short wobble over the last two weeks and turn that into a long term trend.” (41:30)

Where will value accrue in the AI value chain?

• Holding the workflow layer is key—makes it easier to switch models as needed (i.e., commoditization of the model layer is looming).
  • “If you own the workflow, that’s great because you can always swap out the model.” (42:30)
  • Large enterprises, e.g., HSBC, are making deals with open source model providers.

How should “middling powers” approach sovereignty?

• Build at least a “sovereign stack,” even if true sovereignty is elusive.
• Regional alliances and sharing resources are necessary steps.

Could cybersecurity risks halt AI’s growth?

• Azeem notes every technology wave has security risks, but defenses are evolving alongside threats.
  • “But the seriousness... of what these systems, as they get progressively more agentic, can do ... is getting progressively more serious. But ... we don’t hear are the mechanisms of defense that are being built up...” (48:02)

Notable Quotes & Memorable Moments

• On Absorption:
  “API access is the easy bit. It’s the institutional metabolism that is quite hard, very hard in some cases.” (05:14)
• On the Energy Squeeze:
  “You can have a data center but you can’t power it up.” (08:26)
• On Inevitability of Learning by Doing in AI:
  “The only way you’re going to learn how to ... change the way teams work ... is by actually building that capability, by building it yourself.” (36:08)
• On Public Attitudes:
  “It’s an uncomfortable place to be ... it does introduce the idea of a legitimacy crunch.” (29:36)

Concluding Takeaways

• The bottlenecks in AI scaling now lie in energy, compute, and society—not models or datasets.
• Over the next 24 months, the key story will be how well firms, infrastructure, governments, and societies absorb and legitimize the spread of AI.
• Early movers will have an organizational advantage; waiting brings risk of falling behind.
• Structural, economic, and political uncertainties will continue, but efficiency gains and adaptation will reshape the landscape.
• Societal resistance—and how companies communicate their impact—may become the top bottleneck of all.