I recently came across an extended discussion featuring Gavin Baker — one of the most credible and closely followed voices on semiconductor and technology investing — in conversation with Jess, Senior Partner and Head of AI Strategy at Blackstone. Two things stopped us in our tracks. The first was Baker's observation that just 10 basis points — one tenth of one percent — of the global population is currently using AI models in any meaningful way, and that we are already in a severe and worsening shortage of the physical infrastructure required to support that demand. The second was his framing of what happens next: if even 5% of the world's population begins to engage with AI at the level current frontier users do, the resource requirements become, in his word, unimaginable.
That single data point — 0.10% of users, and already the grid is straining — aligns precisely with the thesis we have been building at Monard: that power and water are not incidental requirements of the AI economy. They are its foundation.
The question everyone in finance, technology and geopolitics is asking is deceptively simple: who is winning the race in artificial intelligence? The more instructive answer, offered by some of the most credible voices tracking this industry, is that the race as most people imagine it has not yet truly begun.
Gavin Baker — who managed $17 billion at Fidelity, beat 99% of his peers across two full technology cycles, and now runs Atreides Management — offered a sweeping and often counterintuitive read of where the AI build-out stands today. The picture that emerged was one of profound scarcity sitting just beneath the surface of breathless software progress.
The Bottleneck: This Is Not a Software War — It Is a Physics Problem
The first and most important point Baker makes is that the race for AI supremacy is not being run on lines of code. It is being run on wafers, watts, and water. The leading edge of AI capability is constrained by the physical capacity of Taiwan Semiconductor to produce advanced chips and by the ability of electrical grids to power the data centres that run them.
TSMC's leadership, Baker argues, are the most consequential people in Taiwan. They view themselves as guardians of a legacy built across decades and they are deliberately pacing capacity expansion rather than flooding the market. Jensen Huang of Nvidia visits every three months pushing for capacity to double or triple. The answer is no. The result, in Baker's view, is that the semiconductor industry's most senior figures are inadvertently protecting the global technology sector from the kind of speculative bubble that destroyed value after every prior technology wave.
— Gavin Baker, Atreides Management
Power and Water: The Infrastructure AI Cannot Exist Without
At 10 basis points of global AI adoption, we are already consuming power and water at rates that are straining grids and cooling infrastructure across the United States, Europe and Asia. Data centres are the fastest-growing source of electricity demand in most developed markets. Water consumption for cooling is drawing scrutiny from regulators and communities alike. And this is before the demand curve has meaningfully begun.
Baker's longer-term solutions — orbital compute powered by solar panels in space, and nuclear energy as the only baseload source capable of matching AI's appetite — are credible on a 5 to 15 year horizon. But those solutions do not address the demand that exists today, or the demand that will arrive in the next three to five years. That gap — between the grid as it stands and the power and water that data centres need right now — is where Monard operates.
The Monard Response: Behind the Meter, Delivered as a Utility
Monard's view is that the immediate and medium-term power and water shortfall facing data centre owners will not be solved by waiting for the grid to catch up. It will be solved by behind-the-meter, utility-style solutions that deliver reliable power and process water directly to the facility — without requiring the data centre owner to commit capital to generation or water infrastructure.
We approach data centre owners the way a water utility or electricity distributor approaches an industrial customer: we own, fund and operate the power and water assets, and the data centre operator pays for what they consume. No capital outlay. No infrastructure risk. No grid dependency. Just a complete, contracted utility service — power and water — available at the meter.
The window between now and the point at which nuclear and orbital compute take meaningful load is not a gap to be managed. It is the defining investment opportunity of this infrastructure cycle. Monard is positioned to fill it.
The Chip Race: Nvidia Leads — But Amazon Is Most Underestimated
On chips, Baker is direct: Nvidia remains dominant, but the most mispriced opportunity in semiconductor investing right now is Amazon's Trainium. The reason is architectural. Modern frontier AI models require a switched scale-up network to run inference efficiently. At the time of this discussion, only two functioning versions existed commercially: the one powering Nvidia's GPUs and the one powering Amazon's Trainium. Google's TPU, despite the company having invented the benchmark by which AI chips are judged, made conservative design choices and will not submit its own product to that benchmark.
The Money: Usage-Based Pricing and the Path to $200 Billion
On monetisation, Baker points to a fundamental pricing transition. The era of flat-fee AI subscriptions is ending. Frontier providers are moving to usage-based pricing where the most powerful capabilities are reserved for enterprise customers paying per token. The analogy is the mobile era — consumers bought a fixed allocation of minutes, then paid overage when they exceeded it. The industry discovered people would pay far more than expected. AI is following the same arc. Coding has emerged as the killer application, and Baker believes the act of writing code may itself be the most direct path to artificial general intelligence.
What's Around the Corner: Too Much Growth, Not Enough Power or Water
We are at the very beginning of a demand curve that is unlike anything the infrastructure world has been asked to absorb before. One tenth of one percent of the global population is using AI today in a meaningful way. The grids are already straining. The water tables are already under pressure. The planning systems were not designed for this.
This is not a technology story. It is an infrastructure story. The winners of the AI era will not be determined solely by who writes the best model or ships the most compelling product. They will be determined, in large part, by who controls the foundational resources — reliable power, available water, permitted land — that make intelligence at scale physically possible.
The race in AI has not yet truly begun. The infrastructure race that will determine whether it can be run already has. We intend to be at the centre of it.
The views expressed in this publication represent the internal perspectives of Monard Infrastructure Inc. and are intended solely for informational purposes. Nothing contained herein constitutes financial, investment, legal, or professional advice of any kind, nor should it be construed as such or relied upon when making any investment or business decisions. Past performance is not indicative of future results. Recipients are encouraged to seek independent professional advice tailored to their specific circumstances before acting on any information contained herein.
This briefing synthesises commentary from Gavin Baker (Managing Partner, Atreides Management) in conversation with Jess (Senior Partner & Head of AI Strategy, Blackstone) at an investment industry conference. Views expressed by Baker and Jess are those of the original speakers, synthesised for this note.