Marc Puricelli — Blog

We Need Public Ownership of AI Infrastructure

Tue, 08 Sep 2026 00:00:00 +0000

I couldn’t resist writing about another of Dwarkesh’s conversations with his good friend and AI industry guru Dylan Patel from SemiAnalysis. According to Dylan, after initially losing money on inference, AI labs now generate revenue well above their compute costs. He puts Anthropic’s annual revenue as high as $50 million per megawatt (MW), against compute costs of $10–15 million. Meanwhile, bringing more capacity online remains constrained by supply-chain bottlenecks and power availability. By the end of 2027, he believes annual revenue could reach $70–80 million per MW across Anthropic or OpenAI, depending on model performance and whether they can keep releasing better models.

This explains why everyone is piling money into new compute capacity. The value of each MW is going up dramatically, promising incredible returns for the foreseeable future. The two leading labs produce the most revenue per unit of compute, which allows them to outbid other buyers for new capacity. As stated in the title of Dwarkesh’s post, two private companies could soon control a majority of the world’s compute.

This hoarding of infrastructure is an especially big problem in AI, as access to compute determines how much research and training you can do on future models. We risk a self-reinforcing cycle that would leave these two giants controlling the world’s most advanced AI and almost all the hardware capable of running it.

Putting aside the debate on the AI financial bubble and the wisdom of such large investments, what’s at stake here isn’t merely attractive returns for investors but raw political and economic power. Whether or not current investments are recouped, money and wealth will always flow to those with power. This is the real bet being made: control of the AI frontier will become an essential pillar of power in the 21st century.

As discussed towards the end of the episode, there is a risk of an unacceptable concentration of power in the hands of just a couple of companies if — and it’s a big if — AI turns out to be able to replace the majority of human work in the near future. We’ve recently seen exciting releases of open-weight models that come close to the frontier at affordable prices. But where would these models run at scale? Where would the next generation be trained if the necessary hardware were owned or leased by companies with a vested interest in staying ahead? Even China faces a substantial gap: Dylan estimates that less than 10% of new AI compute is being deployed there, compared with around 70% in the US.

This is why our governments need to own or co-own a considerable share of the world’s new compute capacity. As in other strategic sectors of national importance, such as nuclear power, weapons manufacturing or space exploration, they need a real seat at the table to ensure AI is developed and deployed in a democratic and socially responsible manner, with access extending beyond Anthropic and OpenAI’s customers. It’s also a matter of national security and sovereignty, as AI becomes an increasingly important component of warfare and defence against cyberattacks.

My concern is that legislation is only effective if you have real leverage over the companies you are trying to regulate. In Europe’s case, we are making it harder for our own companies to compete and adapt without affecting how AI develops in the US or China. I recommend Wolfgang Münchau’s excellent piece on the subject.

Taxation is another valid option. There is an interesting paper worth reading on taxing tokens at the point of usage and how such a system could work. Redistributing some of the value created by AI across society will probably soon become necessary, but taxation alone would not address the concentration of power.

The problem, of course, is that most Western governments are in no financial position to make large infrastructure investments as in the past. Public sentiment in Europe is also largely sceptical about AI. This makes investment at the required scale unlikely. If countries miss this opportunity, they may never catch up with the private sector or geopolitical rivals, as greater access to compute helps the leaders develop better models and extend their advantage.

One financing option would be “AI bonds”: investors would lend to a publicly owned venture and receive interest, while ownership remained public. Why not? If Google and Meta can issue debt to finance new data centres, why couldn’t governments do the same? Public-private joint ventures could also bring in capital and technical expertise, provided the public partner retained meaningful control over how capacity was allocated.

Part of this capacity would be leased at market rates to the highest bidders, including Anthropic or OpenAI, allowing local consumers to continue using their models. Another part would be reserved for competing labs and public research, giving them a shot at training large models of their own.

Revenue would cover operating costs, repay lenders with interest, and help fund expansion and replace ageing hardware — AI chips don’t age well!

The real value for society would come from public ownership and control of infrastructure critical to national prosperity and security. It would give us a say in who uses it and who benefits, helping preserve the social democracies we hold dear.

Compute Will Get More Expensive. How Does This Affect Companies Designing Their AI Strategy?

Thu, 13 Aug 2026 00:00:00 +0000

I’ve written previously about the evolution of AI pricing in the short to medium term, and my worry that many companies in a hurry to adopt the technology at today’s prices will find themselves in a bind for a couple of years: they will have encouraged their employees to rely on AI as much as possible, only to see the cost of that usage skyrocket.

I’ve been following this topic closely, so I enjoyed reading Dwarkesh Patel’s 29 July piece on the risk that the price of compute could increase tenfold in the coming years.

As I’ve highlighted in my previous posts, I agree with Dwarkesh’s take that the hard constraints on hyperscalers’ ability to add new compute capacity every year, despite massive investments, mean that demand will soon outstrip supply. Based on the latest earnings calls from Microsoft, Alphabet and others, this is already the case. These constraints are likely to persist for at least a few years: new fabs alone will not solve shortages in memory, power, data-centre capacity, advanced lithography equipment and leading-edge wafer allocation. (See the excellent work by SemiAnalysis on this.)

I struggle, however, to understand Ed Zitron, who argues that hyperscaler spending is fuelling a huge bubble ready to burst. I don’t dismiss the possibility that the amount of capital being deployed is excessive, and that investors’ expectations of future returns are unrealistic. But the investment in data centres and GPUs is justified. The demand is already here — I see it in my day-to-day work and among friends — and will continue to increase rapidly. This infrastructure takes years to plan and build, and the spending needs to happen now.

The returns might still be underwhelming relative to the sums invested, and much of this demand depends on Anthropic and OpenAI’s continued success. AI labs and hyperscalers are clearly betting that they will be able to charge much more in the future, which is one of the oldest business strategies in the world: get your customers hooked on your product and then jack up the prices. Customers may eventually be able to switch models, but those who own the compute infrastructure will have the market cornered.

As highlighted by Dwarkesh, the cost of renting an H100 GPU for a year is still around fifteen times cheaper than employing a software engineer. So there is still plenty of room for AI labs or hyperscalers to increase prices and find customers willing to pay, especially as models get closer to replacing some forms of human work one-for-one.

Patel then argues that the increased cost can be captured as profit either by the AI labs, which can use the lead of their frontier models over competitors as a differentiator, or by the hyperscalers and chipmakers, which will be able to charge more for compute. A combination of both is likely, but I’d argue that Anthropic and OpenAI are in the weaker position. We’ve seen significant competition from smaller labs, including open-weight models — especially out of China — that can come very close to their best.

Politics might be a factor here. It wouldn’t surprise me if the US government discouraged or prohibited US-based companies from relying too heavily on Chinese models, even if they are open-weight. However, Microsoft, Meta and others would still be happy to offer their own proprietary or open-weight models to customers looking for a cheaper alternative to OpenAI or Anthropic.

So where does this leave companies planning their AI strategy and adopting these tools across the board? I don’t think that the risk of price increases is a reason to stop or even slow down current adoption efforts. The biggest barrier to seeing a real ROI from AI deployment in most organisations is the time required to transform the culture, workflows, governance and systems they need to analyse information, make decisions and deliver to customers at the speed AI can enable. This will take years, especially for large organisations. Any business that achieves this transformation before its competitors will find itself in a great position to deploy and reap significant productivity boosts from future models, once the price of compute decreases.

In the meantime, companies should make sure they can precisely measure and compare the ROI of current use cases under consideration. Counterintuitively, this doesn’t necessarily mean relying only on cheaper models, as frontier models might use fewer tokens to achieve the same result, or cost more but produce a significantly better one. This is where the complexity lies: ROI will be measured differently for every company and use case, depending on factors such as the impact on the bottom line, customer experience or satisfaction, time saved and the cost of compute.

I would also focus on using AI to automate high-volume, low-value tasks, as those will rely most of all on the quality of the harness and internal setup. This should make it easier to switch to cheaper or even locally hosted models if necessary.

Finally, a few years of expensive compute might temporarily slow down some businesses’ eagerness to replace part of their headcount with AI. This might buy HR teams and workers a bit more time to reinvent certain jobs, understand where the human touch still matters, and work out how humans and machines should complement each other.

AI Pricing — A More Nuanced Take

Thu, 11 Jun 2026 00:00:00 +0000

In my first post on this blog, AI Pricing Uncertainty, I worried that companies would deploy and quickly become reliant on AI before fully understanding how the pricing of this technology would evolve. The natural assumption is that prices will fall rapidly, and Silicon Valley has made plenty of noise to push companies and users to adopt AI as fast as possible (surprising, I know).

But things are moving quickly, and the simple pricing structure we've known since the start of the LLM revolution is about to get more complicated. Consumer subscriptions and API access to frontier models remain affordable considering what these tools can produce, but prices have been going up — especially API costs, with both Anthropic and OpenAI aligning their enterprise plans with API token rates this spring. For this reason, and a series of other factors I will cover below, I want to revise my initial prediction with a more nuanced take.

The hard constraints on the hyperscalers are well known and still apply: energy, and TSMC's capacity to produce high-performance chips for everyone at once. There are others too — memory shortages, political resistance to new data centres and to AI generally, and the geopolitical sword of Damocles — any of which could significantly slow the arrival of new capacity.

On the demand side, hyperscalers and investors are betting — and I agree — that we're on the verge of an explosion in demand, as AI's most powerful capabilities are revealed to the public once it has access to a person's or an organisation's specific data and context. iPhone users will get their first real taste when Siri can finally connect messages, email, calendar, and the web, and deliver something personalised. Once that happens, AI will become indispensable to most people, as it already is for some of us. The same will happen for businesses when providers like Microsoft make it easy to deploy agents that can access an organisation's data wherever and in whatever format it is stored, and operate securely within a container. Demand will grow at a rapid pace, probably for decades, as AI is integrated into every sector of the economy — drawing on ever larger sets of data and consuming more tokens in the process.

What I underappreciated in my first post is the tremendous incentive to solve this collision between booming demand and constrained supply. Software improvements have already cut the cost per token — routing queries to the cheapest model that can handle them, for a start — and no doubt many more such optimisations remain to be found. On the hardware side, chips designed specifically for inference, such as Google's TPUv7, are improving data centre efficiency further. Most intriguing of all, Huawei has announced an approach to chip-making that forgoes ever-smaller transistors without sacrificing performance. Even if Huawei cannot deliver on its claims, the point stands: innovation is constant, and it can change the equation at a moment's notice.

The model landscape is shifting too. Not every new model is built to run at the frontier: some run locally, others deliver near-frontier performance at a much lower price. And edge computing is coming back in a big way, preparing for a future where day-to-day tasks and simple operations run free on our laptops and phones.

All of this changes how I think AI pricing will evolve for companies. Because AI is a near-direct byproduct of energy consumption, and the infrastructure that produces it will remain constrained for the foreseeable future, there will be pressure to optimise its use the way we optimise fuel today. Suppliers will be incentivised to differentiate their best models from the rest, and will be able to charge more — perhaps much more — for them if they can prove superior output. Efficiency gains will put downward pressure on frontier prices too — but the frontier is, by definition, whatever runs at the limit of available compute, and so far demand has absorbed every gain the moment it appeared. A business might consider paying top dollar for such models for consulting work, strategy, or other critical tasks. For day-to-day operations, I expect mid-tier models to become commodities, competing on price while offering similar performance, used for cloud operations and anything else that cannot be done on device. In that segment of the market, the durable advantage for suppliers won't be the model at all: it will be the product integration (the "harness"), the ecosystem, and the lock-in that comes with both.

Commodity will not mean interchangeable, though. Getting these models to generate real value will require considerable training and tuning inside each company, and deep embedding into the organisation's data ecosystem — work that makes a deployment very sticky even when the underlying model hardly matters. And that customisation will be a large source of revenue in its own right. SAP built an empire on endless customisation; AI suppliers will do the same.

So here is the revision. My first post argued that prices would rise for the next couple of years, until enough capacity comes online — not indefinitely. I still think that will be the case, but prices won't move in one direction: they will stratify. Frontier capability will get more expensive, the middle will become a commodity (though a sticky one, as covered above), and the edge will cost what it costs to operate.

Two main risks remain, however. The first is competitive: cash-rich players will be able to pay for frontier models in the areas where the raw intelligence of the model matters — strategy, research, complex analysis — and the advantage they buy there could raise uncomfortable questions for competition in our market economies, producing even more winner-takes-all dynamics than the mobile and Web 2.0 era did. The second is about control: companies will need to keep their internal knowledge and IP from leaking into the models they train, and preserve the ability to walk away from an AI supplier without losing the capacity to operate. There will probably be a plethora of good models at different price tiers to choose from. But the dependency that builds up around their implementation — and around the suppliers behind them — is what organisations must think about, and guard against.

Building Power Apps with Claude

Tue, 02 Jun 2026 00:00:00 +0000

I've built plenty of Power Apps over the years, for two reasons. It's a solid low-to-no-code platform for learning how to build apps, and it lives inside the MS365 ecosystem — so you can wire an app straight into Active Directory, Outlook, calendars, and more, with a bit of help from Power Automate. A simple one takes a couple of hours. Because we don't use MS Fabric at my company (and I don't have Azure access), I've relied mostly on SharePoint lists on the backend — which is almost always the most tedious part of the job.

Building my iOS app (Thames Foot Tunnels, have a look) — which can now be done barely having to open Xcode — got me wondering whether the same was possible with Power Apps.

I tried a few approaches with Claude: the pac CLI / Microsoft Power Platform CLI, or exporting a blank app, having Claude modify the .msapp package, and re-uploading it. All hit blockers. Then today, a bit of digging turned up a Microsoft Learn page on a new skill that lets you co-author canvas apps with AI. I pointed Claude at it and, presto: a full World Cup Pick app, built in 20–30 minutes. And all of this from a Mac!

What a time to be building apps.

AI and the Myth of Automatic Progress

Wed, 29 Apr 2026 00:00:00 +0000

I really admire Gary Stevenson and his YouTube channel Gary's Economics. I find it incredible that an ex–City trader has become such an influential voice in the UK political ecosystem in little more than a year, largely through the effectiveness of his videos. His latest video is excellent and, in my view, one of the most convincing analyses of the broad societal impacts AI is likely to have.

Although Gary studied economics at some of the best schools in the country, he has observed first-hand, through his work as a trader, that many mainstream economic models struggle to explain real market behaviour. Instead, what he finds most revealing is how these economic narratives are used to justify government policies and maintain the status quo — more than what they claim as a "science" of economics.

In this latest episode, he takes on the common belief that technological progress is always — or at least in the long run — beneficial to humanity. He argues that this assumption rests on a naïve and anachronistic reading of the Industrial Revolution. While it is true that the surge in production capacity since that period has dramatically raised living standards globally, the elephant in the room is that it took nearly 200 years of hardship for most people to benefit.

Industrial workers, including children, often worked 12-hour days, six days a week, in harsh conditions. At the same time, large numbers of skilled workers — such as artisans in places like India — lost their livelihoods as they were unable to compete with mass-produced British goods, which they were often forced to consume. The list goes on. Generations had to struggle — and in many cases fight — for the gains of increased production to be shared more broadly.

Gary's point is simple but important: technological progress does not automatically translate into better lives for the majority, even over time. If we do not fix what is already broken, AI risks accelerating and amplifying existing inequalities. There is also the possibility of more extreme outcomes, including forms of mass surveillance or automated violence. Preventing that is a matter of choice.

I consider myself an AI optimist. I believe AI has the potential to enable humanity to achieve extraordinary things, at scales far beyond what we can do today. However, it is clear that those of us involved in building and deploying these systems have a responsibility to engage with these questions — and to push companies and governments to make choices that ensure AI, and the infrastructure behind it, improves the lives of the many—not just the few.

AI Pricing Uncertainty

Wed, 22 Apr 2026 00:00:00 +0000

Companies are rushing to deploy AI as fast as possible. Large, well-established businesses fear being left behind by competitors or replaced by new, agile, AI-first startups. This is understandable — it's what happened in the last shift to e-commerce, mobile, and social media. But there's another cost — besides the missed opportunity cost — that they should be mindful of: the price of AI itself.

It's well known that Silicon Valley's playbook is to launch products at low, subsidised costs to gain market dominance before increasing prices. The current pricing of AI subscriptions is extremely affordable, especially for the latest models, which are very capable. However, the cost per million tokens looks set to increase in the coming years. Despite the frankly ridiculous levels of investment in AI infrastructure in recent years, many analysts¹ are warning that it might not be enough to meet rapidly growing demand. Supply will be constrained by chip production capacity and energy availability, both of which take time to ramp up. The recent troubles in the Middle East could delay this even further. And investors will eventually demand a return on their investments, putting pressure on AI suppliers to raise revenue. We're already seeing Anthropic struggling to meet demand and raising prices.

Every organisation figuring out its AI strategy should pay attention to the risk of rapidly increasing costs. As pointed out by Ben Thompson of Stratechery, AI is fundamentally different from SaaS because the latter has a marginal cost of zero, which resulted in predictable costs and declining per-user prices as companies scaled. AI is the opposite: it costs more the more you use it. Output quality is also correlated with the price users are willing to pay — better models, running on better hardware and using more tokens, produce better results. Organisations adopting AI will not all be on equal footing, as larger and better-resourced players will be able to afford superior models or use them more extensively, giving them a competitive advantage. This could result in a bidding war among cash-rich competitors, pushing prices even higher and leaving everyone else behind.

Furthermore, it's unclear how quickly they can expect cost savings, productivity gains, or revenue growth from their AI investments. Recent studies have shown that few have been able to implement AI in ways that produce meaningful results. It's becoming clear that doing this well requires more than adding AI chatbots to existing software (hello, Copilot!). It requires a complete rethink of workflows, business processes, and certain jobs. Substantial backend work will be necessary to build the data pipelines, context-awareness, and security guardrails for LLMs to complete tasks end to end, correctly. This will take years and cost money. Until that transformation is complete, AI gains will remain limited and organisations will not be able to reduce their workforce — with the exception of specific jobs already ripe for replacement, like certain developer and customer-care roles. This period will therefore require heavy investment in AI infrastructure and workforce education while maintaining the same operational costs. AI could result in a significant increase in operational expenses for years before any ROI is gained.

Despite all of these risks, it's clear that the potential of AI to reduce costs and increase productivity is too promising to ignore, and refusing to embrace technological change is most likely a death sentence over time. But incumbents should beware of complete AI dependency, especially given the pricing uncertainty ahead. It is a risk to assume that the cost of AI — especially for the best-performing models — will trend downward. Once that transformation is complete, there will be no going back, and bargaining power will be squarely in the hands of suppliers. It's also unclear how much savings on salaries and benefits can offset AI costs — many new types of jobs will likely be created, and jobs that cannot be automated will become more expensive. SaaS has shown that system lock-in is real. It can be extremely painful and costly to change software providers. This could become much worse when an entire organisation depends on one deeply integrated AI layer.

Obviously, it's possible all of this will be fine, and that markets and societies will gracefully adapt, as they historically tend to. But incumbents should plan for contingencies to ensure they can reduce their dependency on AI suppliers — the same way they manage supply chain risks today. Ultimately, it's crucial that the real value engine of a company remains its IP and its people, not the AI layer it happens to be using.