2025-11-30 12:30:46
Hi all,
This is the final post in our four‑part series, Ten things I’m thinking about AI. It maps the next 24 months of AI as we mark ChatGPT’s third anniversary – and it’s the most comprehensive macro take we’ve published this year.
If you missed the first three parts, catch up here:
Part I: The Firm [read here, no paywall]
Part II: Physical limitations [read here]
Part III: The economic engine [read here]
Part IV: The macro view [today’s email]
Sovereign AI is splintering the stack
A growing utility-trust gap risks a societal schism
In early 2025, sovereign AI was mostly posture. The AI landscape has become starkly bipolar: the US controls 75% of global AI compute, China 15%, leaving everyone else marginalised. You can adopt US/Chinese AI systems and become vulnerable to data exploitation, service restrictions, embedded values, and unfavourable terms. Or accept weakness, limit adoption and fall behind as frontier states achieve economic, scientific, and military breakthroughs.
You can see the logic hardening. Britain is building national infrastructure with Nvidia; the Gulf is spinning up its own polity via the $100 billion MGX vehicle. OpenAI’s “Stargate” initiative fits this fractured landscape perfectly. Stargate UAE offers a gigawatt-scale compute cluster under US-aligned governance. Meanwhile, South Korea joins the orbit as a chip arsenal, and Saudi Arabia finances “AI factories” for local workloads.
Even Brussels is softening regulations to keep champions like Mistral onshore. Officials now discuss compute and models as strategic assets to be rationed through alliances, not left to the market.
The challenges are most acute for mid-sized and smaller nations, where mid-sized means anything smaller than the US or China. One possible route may be multinational or minilateral approaches that involve pooling capabilities or resources, as argued in this working paper from an Oxford University research group. This question is also the focus of a working group I co-chair on how governments need to approach next-generation computing infrastructure.
This logic virtually guarantees a splintered world. We are moving toward distinct US-aligned, China-aligned, and “non-aligned” stacks. The era of the borderless internet has come to an end; the era of the sovereign stack has begun.
We are living through a massive psychological split. On the one hand, utility is winning: 900 million people now use ChatGPT and other LLMs to code, write and think, bringing the “sovereign stack” into their personal lives. The utility is too high to ignore.1 Yet, familiarity has bred fear, not contempt.
Three-quarters of Americans now believe AI will negatively impact their lives, a figure that has hardened even as adoption skyrockets. We have become pragmatic addicts, dependent on a tool we fundamentally do not trust. This isn’t a “tech backlash” of the previous eras. I see it as a new kind of social bargain – we will use the machine to do our jobs today, even while we suspect it will take our jobs tomorrow.
2025-11-30 11:00:55
Hi all,
ChatGPT launched three years ago on this day. In that short time, access to “good enough” intelligence has grown massively.
I’ve spent a decade analysing the exponential age (I wrote a book about it in 2021), but these three years mark a distinct turn with overlapping S-curves in models and chips, new behaviours, and a growing gap between exponential technology and linear institutions.
A four-part series sets out my current view of AI as a general-purpose technology at this inflection point. I cover how I see it reshaping firms, markets and the wider economy:
Part 4 will be in your inbox today, right after the Sunday edition!
Anthropic’s new study says that even if AI progress stopped today, existing tools could lift US labor productivity by about 1.8% a year for the next decade – a historically unprecedented pace compared with the post‑war boom or the dot‑com era. I agree with this sentiment – in fact, I think it’s probably been true since GPT-4. The tools are powerful but implementation is tougher: companies need to go through wide-scale people-centric transformation processes to make best use of them. This takes time.
Anthropic’s projection matches what many economists expect. But we’re not seeing it yet: recent US productivity growth is around 1.5%, roughly the historical average.
Although visible AI use is clustered in tech hubs and roles like software development, much larger “submerged” exposure exists across admin, finance, and professional services and makes up 11.7% of wage value, five times higher than the visible tech sector (see MIT’s Project Iceberg for details on this research). For this reason, the biggest impacts will emerge outside classic tech centers.
So far, the impact isn’t showing up clearly in the broad stats – and it’ll matter that it does in 2026.
📈 Quick read: Not a bubble yet. Application-layer revenue is strengthening, but funding quality continues to show signs of strain.
Three key movers this week:
Nvidia has issued a detailed memo to analysts systematically refuting bear cases regarding circular financing and accounting irregularities. It came across as a little desperate.
2025-11-29 14:08:41
Hi all,
This is Part III of a four‑part, members‑only series Ten things I’m thinking about AI as we approach the third anniversary of ChatGPT.
The series connects the dots with a view to the past and to the future. From energy and depreciation schedules to capital markets and platform strategy, it’s a unified map for the next 24 months of AI – grounded, contrarian where needed, and practical. Here’s what we cover:
Part I: The Firm [read here, no paywall]
Part II: Physical limitations [read here]
Part III: The economic engine [today’s email]
Capital markets struggle with exponentials
Perhaps GPUs do last six years
Compute capacity is existential for companies
The “productivity clock” rings around 2026
Part IV: The macro view [coming up]
2025-11-29 01:31:28
In today’s session, I used ChatGPT’s third birthday as an opportunity to show how the exponential age is unfolding in real time. The technology improves rapidly, people integrate it into their lives but our institutions adjust much more slowly - a tension I explored in my first book.
Enjoy!
Azeem
2025-11-28 22:09:20
Hi all,
Ahead of ChatGPT’s 3rd birthday on Sunday, I’ve put together a four-part series on my current beliefs about AI, in which I’ll cover:
Part I: The Firm [read here, no paywall]
Part II: Physical limitations [today’s email]
Part III: The economic engine [coming up]
Part IV: The macro view [coming up]
That simple “exchange” with Gemini, if done by all 4.5 billion people who use Google, would use 1-1.5 gigawatt‑hours of electricity, roughly 20 minutes of London’s power demand.
And that is just the baseline. Some in our team occasionally consume 50 million tokens a day. If 4.5 billion people reached that level of use, it would exceed global annual electricity consumption.1
This highlights the massive compute and energy load we are potentially dealing with in AI. Is there enough raw capacity to meet that demand?
Microsoft, Amazon and Google have all highlighted that extreme compute demand is causing bottlenecks. But the compute is physical: it needs to be housed in datacenters, filled with racks, cooled and powered by electricity.
At some point, you run up against physical bottlenecks. Some of this is the demand for the chips, processing and memory, evidenced by Nvidia’s reported $500 billion order backlog and the fact that SK Hynix has officially booked out its entire high-bandwidth memory (HBM) capacity through 2026. While silicon availability is one bottleneck, it is not the ultimate one.
Energy is the most significant physical constraint on the AI build-out in the US, as I argued in the New York Times back in December. The lead time for new power generation and grid upgrades, often measured in decades, far exceeds the 18-24 months needed to build a data center. The US interconnection queue has a median wait of four to five years for renewable and storage projects to connect to the grid. Some markets report average waiting times as long as 9.2 years.
This is also a problem in Europe. Grid connections face a backlog of seven to ten years in data center hotspots.
For the Chinese, the calculus is different. points out that “current forecasts through 2030 suggest that China will only need AI-related power equal to 1-5% of the power it added over the past five years, while for the US that figure is 50-70%.”
Because the American grid can’t keep up, data center builders are increasingly opting out and looking for behind-the-meter solutions, such as gas turbines or off-grid solar. Solar is particularly attractive – some Virginia projects can move from land-use approval to commercial operation in only 18 to 24 months. Compute will increasingly be dictated by the availability of stranded energy and the resilience of local grids rather than by proximity to the end user.
These grid limitations cast doubt on the industry’s most ambitious timelines. Last year, some forecasts anticipated 10 GW clusters by 2027. This now appears improbable.
2025-11-27 02:41:21
On Sunday, 30th November, ChatGPT turns three. Wild three years.
It’s triggered a dramatic race in research. Before ChatGPT launched, the big US labs released a major model every six months. After ChatGPT, releases picked up to one every two weeks.
It’s created a cascade of spending commitments. Big tech firms have increased their capex from about $120-150 billion per year before ChatGPT, to closer to $400 billion this year. We reckon that in 2025, about $220-230 billion of that is incremental investment to meet AI demand. The bankers are struggling to make sense of it all.
As ChatGPT approaches its third birthday, I want to summarise my current beliefs about the deployment of the technology and the wider environment. I first shared these with members on the EV member Slack.
I will go into more detail in a four-part series over the following days, to cover:
Part I: The Firm
Enterprise adoption – hard but accelerating
The revenue rocket
Part II: Physical limitations
Energy constraints limiting scaling
The inference-training trade-off
Part III: The economic engine [today’s email]
Capital markets struggle with exponentials
Perhaps GPUs do last six years
Compute capacity is existential for companies
The “productivity clock” rings around 2026
Part IV: The macro view
Sovereign AI fragments stack
Utility-trust gap dangerously widens
Today’s first part will focus on the firms, looking at how adoption and revenue are materializing. The next three parts will analyze the physical build-out of compute, the ecosystem’s new economics and the wider macro-political system.
Today’s post is open to all – so share widely. Parts 2-4 of the series will be open to paying members of Exponential View.
Here are the 10 things I think about AI right now.
There is a clear disconnect between the accelerating spend on AI infrastructure and the relatively few enterprises reporting transformative results.
By historical standards, the impact is arriving faster than in previous technology waves, like cloud computing, SaaS or electricity. Close to 90% of surveyed organizations now say they use AI in at least one business function.
But organizational integration is hard because it requires more than just API access. AI is a general-purpose technology which ultimately transforms every knowledge-intensive activity, but only after companies pair the technology with the institutional rewiring that’s needed to metabolise them. This requires significant organizational change, process re‑engineering and data governance.
McKinsey shared that some 20% of organizations already report a tangible impact on value creation from genAI. Those companies have done the hard yards fixing processes, tightening data, building skills and should find it easier to scale next year. One such company is BNY Mellon. The bank’s current efficiency gains follow a multi-year restructuring around a “platforms operating model”. Before a single model could be deployed at scale, they had to create an “AI Hub” to standardize data access and digitize core custody workflows. The ROI appeared only after this architectural heavy-lifting was completed. The bank now operates over 100 “digital employees” and has 117 AI solutions in production. They’ve cut unit costs per custody trade by about 5% and per net asset value by 15%. The next 1,000 “digital employees” should be less of a headache.
The best example, though, is JP Morgan, whose boss Jamie Dimon said: “We have shown that for $2 billion of expense, we have about $2 billion of benefit.” This is exactly what we would expect from a productivity J‑curve. With any general‑purpose technology, a small set of early adopters captures gains first, while everyone else is reorienting their processes around the technology. Electricity and information technology followed that pattern; AI is no exception. The difference now is the speed at which the leading edge is moving.
I don’t think this will be a multi‑decadal affair for AI. The rate of successful implementation is higher, and organizations are moving up the learning curve. As we go from “hard” to “less hard” over the next 12-18 months, we should expect an inflection point where value creation rapidly broadens. The plus is that the technology itself will only get better.
Crucially, companies are already spending as if that future value is real. A 2025 survey by Deloitte shows that 85% of organizations increased their AI investment in the past 12 months, and 91% plan to increase it again in the coming year.
One complicating factor in assessing the impact of genAI on firms is the humble mobile phone. Even if their bosses are slow to implement new workflows, employees have already turned to AI – often informally, on personal devices and outside official workflows – which introduces a latent layer of traction inside organisations. This is a confounding factor, and it’s not clear whether this speeds up or slows down enterprise adoption.
On balance, I’d expect this to be the case. In diffusion models inspired by Everett Rogers and popularised by Geoffrey Moore, analysts often treat roughly 15-20% adoption as the point at which a technology begins to cross from early adopters into the early majority1. Once a technology reaches that threshold, adoption typically accelerates as the mainstream follows. We could reasonably expect this share to rise towards 50% over the coming years.
However, 2026 will be a critical check-in. If the industry is still relying on the case studies of JP Morgan, ServiceNow and BNY Mellon rather than a slew of positive productivity celebrations from large American companies, diffusion is taking longer than expected. AI would be well off the pace.
We estimate that the generative AI sector experienced roughly 230% annual revenue growth in 2025, reaching around $60 billion2.
That puts this wave on par with commercialization of cloud, which took only two years to reach $60 billion in revenue3. The PC took nine years; the internet 13 years4.
More strikingly, the growth rate is not yet slowing. In our estimates, the last quarter’s annualized revenue growth was about 214%, close to the overall rate for the year. The sources are familiar – cloud, enterprise/API usage and consumer apps – but the fastest‑growing segment by far is API, which we expect to have grown nearly 300% in 2025 (compared to ~140% for apps and ~50% for cloud). Coding tools are already a $3 billion business against $157 billion in developer salaries, a massive efficiency gap. Cursor reportedly hit $1 billion ARR by late 2025, the fastest SaaS scale-up ever, while GitHub Copilot generates hundreds of millions in recurring revenue (see my conversation with GitHub CEO Thomas Dohmke). These tools are converting labor costs into high-margin software revenue as they evolve from autocomplete to autonomous agents. The current market size is just the beginning.
Consumer revenues, meanwhile, are expanding as the user base compounds. Monthly active users of frontier chatbots are driving a classic “ARPU ratchet”: modest price increases, higher attach rates for add-ons, and a growing share of users paying for premium tiers. There are structural reasons to expect this to continue, even before AI feels ubiquitous inside firms.
First, the base of adoption is widening. If 2026 brings a wave of verified productivity wins, this trajectory will steepen. More firms should enjoy meaningful results and the surveys should show unambiguously that 25-30% of firms that started pilots are scaling them. As the remaining majority shift from pilots to production, they will push a far greater workload onto a small number of model providers. Revenue can rise even while most firms are still doing the unglamorous integration work; pilots “chew” tokens, but scaling up chews more.
Second, the workloads themselves are getting heavier. A basic chatbot turn might involve a few hundred tokens, but agentic workflows that plan, load tools and spawn sub‑agents can consume tens of thousands. To the user it still feels like “one question,” but under the surface, the token bill – and therefore the revenues – is often 10-40x higher.
Of course, this growth in usage will see token bills rise. And companies may increasingly use model-routers to flip workloads to cheaper models (or cheaper hosts) to manage their bills.
But ultimately, what matters here is the amount consumers and firms are willing to spend on genAI products.
Tomorrow, we’ll turn from the demand side to the supply side – and confront the physical constraints that are shaping the industry’s trajectory. Parts 2-4 are available exclusively to paying members of Exponential View.
I tend to view 6% as a cross point.
This is our tightest, conservative model which only looks at deduplicated spend or uplift from genAI services. If a company offers a bundled product with a genAI element, we try to isolate that element. We exclude revenue from services firms.
2025-adjusted.
Assuming “commercial launch” of the internet (for ads) is 1994.
