Reading the room in a room of code. That has become my ritual: four in the morning in Tallinn, the apartment smelling of cold coffee and the low hum of a GPU miner in a closet I still haven't had the heart to unplug. I am not mining anything anymore. I am running Python scripts, cross-referencing trade statistics, power-grid capacity data, and the vocabulary that keeps shifting beneath the industry's feet. The goal is simple: determine which stories are real and which are press-release confetti.
The latest story to cross my screen comes from Crypto Briefing, and its headline does not seem like it should stop anyone in San Francisco — until you sit with it for longer than a scroll. Mexico emerges as key player in US AI infrastructure boom. Slow down. Read it again. The United States, the country that trains the world's most powerful artificial intelligence systems, is now depending on a neighbor best known for avocados, auto plants, and cartel headlines to power its algorithmic future.
Here is the paradox that kept me awake. Mexico does not train frontier models. It does not design GPU architectures. It does not produce the papers that win NeurIPS awards. There is no Mexican OpenAI, no Anthropic del Bajío, no DeepSeek in Guadalajara. Search for “Mexican AI companies” and you will find logistics software, call-center automation, and a few fintechs. And yet, this same country exported roughly $475 billion worth of goods to the United States in 2023, overtaking China as America's largest trading partner. Its border industrial corridors are filling with factories that will build the physical scaffolding of artificial intelligence: server racks, power conversion units, cooling loops, steel enclosures. The name of the game, if you read between the lines, is not intelligence. It is electricity, geography, and the legal architecture that lets one nation supply another with the raw stuff of computation.
What exactly is being exported when the phrase “AI exports” appears in a headline? The article never says. That ambiguity is almost certainly not a failure of reporting. I don’t think it is a failure of terminology either. I think the ambiguity is the entire story. AI has entered an era in which the bottleneck is no longer the model — it is the physical machinery of distribution. Land for data centers. Transmission lines. Cooling water. Tariff schedules. The willingness of one country to keep the lights on for another. That is the room I am reading. It is a room full of watts, not weights.
Let me go deeper into what this actually means, because the shorthand version being peddled to investors misses the point entirely.
To understand why Mexico matters, you have to understand the crisis unfolding inside the American grid. AI data centers are electricity monsters. A single large training cluster — say, one with 100,000 NVIDIA-class accelerators — draws between 600 megawatts and a full gigawatt. Let me put that number in human terms. A gigawatt is roughly the output of an entire nuclear power plant, or a medium-sized city. One warehouse. One cluster. One city's worth of continuous electrical load. Microsoft, Amazon, Google, and Meta are competing for a resource that the United States simply does not have enough of, in the places that they want to build.
The American grid was designed in an era when electricity demand was flat or declining. That era ended. Data centers, electric vehicles, heat pumps, and reshored manufacturing have collided at exactly the same moment. The queue for grid interconnection in the United States now stretches for years; transmission lines take five to ten years to permit; and transformer lead times float somewhere past two years. Hyperscalers have hit what industry veterans call the grid wall.
The consequence is an extraordinary mental shift inside the capital-allocation teams of the world's largest technology companies. They no longer ask, “Where is the best talent?” They ask, “Where is the power, and how fast can we get it?” That is the question that turns a map of North America into a treasure map with a big red X drawn across Mexico.
Mexico checks more boxes than almost any competitor. Land is abundant: vast industrial parks in Monterrey, Chihuahua, and Hermosillo already exist with fiber, rail, customs facilities, and security perimeters. Energy is comparatively cheap: Mexico has roughly 30 gigawatts of installed wind and solar, and industrial electricity prices can fall as low as four to six cents per kilowatt-hour — below most major US hubs. Geography is a competitive weapon: a trucking route from Monterrey to Laredo, Texas, takes less than a day. And the legal framework is already in place: the United States-Mexico-Canada Agreement, known as USMCA, provides duty-free treatment for goods that meet its rules-of-origin requirements.
Then there is the industrial inheritance. For three decades, Mexico has been the workshop of the North American auto industry. That ecosystem — parts suppliers, assembly-line labor, quality-control engineers, export logistics — is exactly the skill base required for a new kind of manufacturing. Tesla is assembling in Nuevo León. Foxconn is expanding in Chihuahua. General Electric has announced new facilities in the border region. The hardware of AI looks different from the hardware of combustion engines, but the discipline of building complex physical products in volume under tight timelines transfers remarkably well.
Now here is the part the press releases leave out: the core of the Mexican opportunity is not labor arbitrage. It is an energy story wearing a manufacturing costume.
Let me walk through the energy arithmetic, because the numbers are not merely interesting — they are the reason this entire narrative exists.
The AI infrastructure boom is, at bottom, an electricity procurement crisis. A training cluster at the scale that frontier labs now contemplate — a hundred thousand accelerators and climbing — presents a power requirement so large that it overloads the planning horizon of most utilities. In the United States, building a new natural-gas combined-cycle plant to serve that load can take longer than the expected lifespan of the GPU generation it is meant to power. This is the fundamental misalignment: the AI industry moves on an eighteen-to-twenty-four-month innovation cycle, but the energy industry moves on a five-to-ten-year infrastructure cycle. Something has to give.
Mexico offers a compressed alternative. The country has significant proven natural-gas reserves and a pipeline network that already connects to Texas. Its northern states have seen an influx of private power generation investment under energy reforms that allowed foreign participation in electricity markets. A developer can realistically bring a new combined-cycle gas plant online in the Mexican north in three to four years — meaningfully faster than equivalent projects in most of the United States. When you add the existing wind and solar capacity in areas like Oaxaca and Baja California, Mexico becomes a diversified energy supplier, not just a fossil-fuel stopgap.
The cost curve does the rest. AI companies live and die by their power curves. A two-cent-per-kilowatt-hour difference across a gigawatt-scale facility translates into tens of millions of dollars per year — permanently, not just a one-time saving. That arithmetic is too large for any CFO to ignore. The financial engineering of AI infrastructure, in other words, increasingly flows through the question of whose grid you plug into.
But wait. Electrical generation is one piece of the puzzle. The other is transmission, and this is the sector where Mexico's role starts to look less like a supplier and more like a structural pivot. US grid operators have announced or begun planning at least five new cross-border transmission line projects between Mexico and the United States. These lines, originally framed as reliability enhancements, are rapidly being reframed as AI-infrastructure enablers. This is the quiet work of the boom: electricity crossing the border in both directions depending on demand, with Mexico serving as a power reservoir for the American Southwest.
I don’t think it is hyperbolic to call this a new energy geography. And geography, once fixed, creates enormous stickiness. Once the transmission lines are built, once the gas plants are running, once the industrial parks are wired for gigawatt-scale loads, Mexico ceases to be an alternative and becomes a necessity. That is the kind of advantage that no amount of algorithmic innovation can replicate.
Now let me shift from raw physics to the legal architecture that makes this physical flow possible. Because none of this happens without the fine print of USMCA.
USMCA was sold to the American public as a victory for autoworkers and dairy farmers. But like all trade architecture, it contains the seeds of uses its drafters never anticipated. The agreement's rules of origin specify what percentage of a product's value must come from North America to qualify for tariff-free treatment. Those rules, calibrated for an age of cars and air conditioners, are now being applied to a different kind of product: AI server racks. And the results are reshaping global supply chains.
Consider a typical GPU server assembly. The chips come from Taiwan, fabricated in facilities that are the most geopolitically strategic factories on Earth. The memory components come from Korea. The high-end printed circuit boards may come from Southeast Asia. None of that, by itself, qualifies as North American content. But when those components are brought into a plant in Monterrey, combined with locally fabricated steel enclosures, locally assembled power distribution units, locally cut cables, and local labor that performs systems integration, testing, and final configuration, the resulting product can satisfy USMCA's regional-value-content threshold. A server rack that was twenty percent “North American” becomes seventy percent. It crosses the border duty-free.
That is the quietly elegant magic of the arrangement. The tariff code, not any national policy called “AI strategy,” is what draws the factory map of the digital age.
The US government is not unaware of this. The CHIPS and Science Act explicitly endorsed a framework called “friend-shoring” — the idea that critical supply chains should be located in politically allied countries. Mexico, by virtue of geography and trade law, is the ultimate friend-shore. The Biden administration's Americas Partnership for Economic Prosperity added further diplomatic scaffolding, framing the region as a unified production block. When I translate this for institutional clients, I put it in the simplest possible terms: the United States is building a North American compute supply chain to hedge against the possibility that its Asian supply chain becomes inaccessible. Mexico is the hedge. And hedges, when the risk materializes, cease to be optional.
Here is the behavioral dimension that my analyst brain can't stop probing. The language of “nearshoring” obscures something more interesting: this is not just firms seeking cheaper inputs. It is firms seeking options. The Mexican AI infrastructure boom is a portfolio play against the uncertainty of the 2020s. By building capacity in Mexico, American hyperscalers buy themselves optionality: the option to shift production closer to home, the option to access power that the US grid cannot deliver, the option to preserve margin in a world of tariffs. This is a psychological story as much as an industrial one. Reading the room in a room of code, I see the same pattern I observed in the NFT mania of 2021 — people were not buying JPEGs, they were buying identity claims. And today, hyperscalers are not buying land, they are buying geopolitical insurance.
Now let me examine what actually gets built on the ground. Because the press release version of the story — “Mexico will manufacture AI servers” — hides a more complex division of labor.
Walk through a modern AI factory and you will notice something immediately: the floor is full of racks, cables, and precision cooling infrastructure, but the heart of the machine is shipped in from elsewhere. A single GPU server rack is a carefully constructed assembly of components from a dozen countries. The GPU modules and networking switches are the crown jewels; they arrive in secure containers with the kind of logistical choreography reserved for military hardware. The chassis, the power supplies, the backplane, the liquid-cooling manifolds — these are the parts that can be, and will be, built in Mexico. They are not trivial. They account for a significant share of the rack's total weight, its thermal performance, and its reliability over a five-year operating life. But they are not the algorithmic soul of the system.
That distinction matters enormously. Mexico's manufacturing role can be described as “final assembly and systems integration.” It is a high-value activity, but it is not the highest-value activity. The margin on a finished AI server rack is not split equally among all participants. The design, the chip ecosystem, and the software standard stack retain the dominant share of economic profit. Mexico captures the manufacturing value, the energy value, and the logistics value. It does not capture the intellectual property value.
This is not entirely a bad thing. Manufacturing value is real, and it has positive externalities that algorithm value does not: jobs, infrastructure, training, ecosystem formation. An engineer who spends five years learning how to balance thermal loads in a liquid-cooled AI rack at a plant in Chihuahua has acquired a skill that current labor markets sorely lack. Over time, ecosystems of this kind compound. But the asymmetry deserves a clear-eyed label. Mexico is becoming the brawn of the AI economy. The brain remains concentrated in California, Seattle, and increasingly, Chinese research centers.
The China question weaves through every layer of this analysis, and it is impossible to avoid. Let me be direct: the Mexican manufacturing boom exists, in significant part, because the United States wants to reduce its dependence on Chinese supply chains. But the same features that make Mexico attractive to American capital — proximity, tariff-free access, low-cost energy — make it attractive to Chinese capital as well. Chinese AI server manufacturers, including names like Inspur, Huawei, and H3C, have well-documented interest in accessing the American market through third countries. Mexico is the most logical third country on the continent. The question is not whether Chinese firms want to insert themselves into the Mexico-America supply chain. The question is whether the United States will permit it.
The obvious regulatory response is expanded export controls. The US Commerce Department has already demonstrated its willingness to restrict the flow of advanced AI hardware, and Mexico's role as a potential transshipment route is the kind of problem that keeps compliance officers awake. I expect new rules targeting the re-export of controlled AI components through Mexican entities within the next twelve to twenty-four months. When that happens, we will learn something important: the Mexican AI boom is not just a story about energy and trade — it is a story about which countries are allowed to touch the machinery of intelligence.
Let me now bring the analysis down from the macro level to the level of the balance sheet. Because the investment narrative around Mexico is, in my view, both the most overhyped and most underexamined part of this story.
There is a recognizable pattern in capital markets when a new infrastructure theme emerges. It happened with fiber optics in the late 1990s, with shale energy in the 2010s, and with data centers more recently. A “theme premium” develops: companies in the same zip code or, in this case, in the same country, see their valuations re-rate upward far faster than any actual cash flows. The Mexican AI theme is exhibiting exactly this behavior. Industrial real estate investment trusts with exposure to Mexican border properties — names like FIBRA Prologis and FIBRA Monterrey — have seen their valuations re-rate on the expectation of AI-driven rental growth. Renewable energy developers, construction firms, and logistics operators in Mexico are all receiving the AI halo.
Some of this premium is justified. The underlying demand signals are genuinely strong. Industrial vacancy rates in northern Mexico have fallen to historic lows. Rental growth in cities like Monterrey has accelerated well above the national average. The appetite of American technology companies for Mexican energy procurement has grown from a curiosity to a documented market. If the AI capital expenditure cycle continues — and the current guidance from Microsoft, Google, and Amazon suggests it will — then the fundamentals will reward many of these assets.
But this is where I sound the cautionary note that is almost always missing from the bullish headlines. The Mexican AI infrastructure theme is a derivative investment. Its value is derived from the capital expenditure decisions of a handful of American firms. If those decisions stall — if a recession hits, if the AI bubble deflates, if a technology innovation makes current infrastructure obsolete — the theme will correct violently. I do not use the word “violently” lightly. A growth narrative that is entirely dependent on external capital flows and external technology cycles has no floor of its own. The peso depreciates, the REITs reprice, and the factories sit half-built.
There is also the question of who actually captures the returns. American enterprises building facilities in Mexico will route profits through American holding companies. The Mexican subsidiaries may generate revenue, but the intellectual property and the financing structures remain elsewhere. The economic benefit to Mexico is real — jobs, wages, tax revenue — but it is smaller and slower than the topline numbers suggest. In my experience translating institutional supply-chain economics into plain language, the distinction between “where the factory is” and “where the value accumulates” is the single most misunderstood concept in the entire nearshoring discussion.
The currency layer adds further texture. A US investor who puts dollars into Mexican infrastructure assets takes on peso-denominated exposure. The peso has been strong in recent years, which boosts dollar returns. But Mexico's currency is sensitive to US political cycles, interest-rate differentials, and the volatility of energy prices. A sudden shift in any of these variables can wipe out a year of compounding gains. This is not a reason to avoid the theme; it is a reason to size positions carefully and understand the full risk surface.
Let me now zoom out to the long arc, because the most important insight about Mexico is not economic, it is structural. I have spent the last year thinking about the idea of the “autonomous economy” — the vision in which AI agents transact, trade, and reallocate capital without direct human intervention. My work on this narrative has taken me deep into the question of what such an economy actually rests on. The answer is humbling: it rests on the same physical layer that every previous economy required. Concrete. Copper. Cables. Current.
AI agents do not care about geopolitics. They care about latency, cost, and reliability. A model that performs an inference in a data center in Querétaro is indistinguishable from one in Dallas — except in the wholesale electricity price. This means the economic geography of AI will be determined, in large part, by the physical infrastructure decisions we make today. Mexico is positioning itself to be the domain where a substantial share of the world's low-latency inference workloads will run. That is not an exaggeration. It is the direction in which every signal points.
And yet, and yet. Here is where I must stretch the narrative in the direction it does not want to go. The article I started with, the Crypto Briefing piece on Mexico, trades in the vocabulary of boom and victory. It never once asks the question that should be asked: what does it mean for a country to be the backup? What does it mean to build your entire economic renaissance around being the fallback option of a superpower?
Bear with me, because this is the contrarian core of the whole story.
The most glaring red flag is the absence of AI sovereignty. Mexico will not train a frontier model in the next decade. It will not even try. The country's universities produce solid engineers but not world-leading AI researchers. Its capital markets do not fund speculative research. Its policy ecosystem has shown no appetite for the kind of moonshot investments that produce national AI champions. The result is a profound dependency: Mexico's AI infrastructure economy is entirely a function of American demand. If that demand shifts, the infrastructure has no local user base to fall back on. This is supplier economics, and supplier economics means pricing power resides downstream.
Then there is water. This is the variable that almost no press release mentions, and it is potentially the bullet that kills the largest projects. Data centers are not just electricity consumers; the traditional liquid cooling approaches consume enormous volumes of water — evaporative cooling towers can draw hundreds of thousands of gallons per day, especially in hot climates. Northern Mexico, where most of the new industrial capacity is being built, suffers from chronic drought and over-allocated water rights. The aquifer beneath Monterrey is already stretched. This is not an infrastructure problem with an easy fix. It constrains site selection to either coastal locations or facilities designed around closed-loop liquid cooling — which adds capital cost and reduces the cost advantage that drew the project in the first place. The water constraint is the kind of unsexy, physical bottleneck that model engineers in Silicon Valley never think about, but that will determine which projects are actually completed.
The grid itself is another source of quiet fragility. The state-owned Federal Electricity Commission (CFE) has a mixed record on reliability. Blackouts and voltage instability have occurred in northern industrial zones, most notably during extreme weather events. An AI training cluster requires seven-by-twenty-four continuous power in the most demanding class. Interruptions are not inconveniences; they are catastrophic financial events. The result is that every large AI facility in Mexico requires redundant power systems — on-site generation, uninterruptible power supplies, battery banks — which raise the capital cost of every megawatt. Again, the cost advantage persists, but it is thinner than the initial arithmetic suggests.
Security is the third unwritten and unavoidable risk. I am not going to engage in sensationalism about cartels; the reality is more mundane and more pervasive than the media narrative. The threat is not television-style violence. It is theft of copper and components, ransom attacks on construction firms, corruption in local permitting, and a general friction tax on every logistics movement. The cost of private security, insurance premiums, and the reputational risk of operating a billion-dollar facility in a region with security incidents is a real and ongoing burden. It does not stop the project. It does not mean the boom is fake. But it takes a meaningful slice of the economic upside and reroutes it to security vendors and insurers.
And then there is the political risk — the one that keeps rerunning in my mind with increasing frequency. USMCA's rules can be revised. The trade agreement will face review, and the American political climate around trade is volatile. Mexico has been a useful friend to the United States in a time of great-power competition. But friends can be scapegoated when domestic politics requires. The example of China's own rise and the US reaction is a warning about how quickly the terms of economic partnership can change when the domestic mood shifts. The risk of a US election cycle importing tariffs on Mexican goods, whether for immigration reasons or political theater, is not zero. In fact, it is high enough that the rational strategy for multinational firms is portfolio diversification — which is precisely what many of them are doing. They are building in Mexico and in Vietnam and in India. Mexico is not the only backup. It is one of several.
The deepest contrarian point, however, is not about Mexico's weaknesses. It is about the meaning of its entire trajectory in the larger history of technology. Taiwan built its semiconductor empire not by being the backup plot of the United States, but by a fifty-year, state-led, relentlessly self-improving investment in its own technological depth. Korea did the same. Mexico's current boom is an external gift, a transfer of production capacity driven by American fears. Such gifts can be squandered. The question is whether Mexico will convert this moment into its own domestic technological ecosystem — whether the engineers trained in Chinese-owned and American-owned factories will spin off their own ventures, whether Mexican capital will accumulate enough to fund homegrown AI companies, whether the state will invest in the regulatory and educational frameworks that turn infrastructure hosting into genuine economic development.
The evidence is mixed. There are encouraging signs: the strengthening of Mexican industrial capacity, the growth of a class of skilled technicians, the potential for Mexican institutions to learn from hosting sophisticated global companies. There are discouraging signs too: the persistence of entrenched corruption, the fragility of the rule of law, the absence of a coherent national industrial policy for AI. The future is genuinely open. Mexico could become the Taiwan of the Western Hemisphere — an indispensable manufacturing partner with its own indigenous innovation. Or it could become the Brazil of the 1970s — a country whose growth spurt was real but whose structural dependencies prevented a breakthrough into sustained, self-directed development.
Which brings me to the question I believe every serious analysis of this topic must eventually confront: what is the shelf life of geopolitical insurance? The boom is being driven by fear of China. But fear is a temporary emotion. As the United States strengthens its domestic manufacturing capacity, as small modular reactors and advanced geothermal technologies come online, as the economics of domestic power generation improve — the urgency behind nearshoring will decline. Mexico's moment is real. But moments pass. The only institutions that outlast moments are those that capture not merely the capital flows but the knowledge flows that accompany them.
I don't know whether Mexico will capture those knowledge flows. I do know that the answer will be written not in research papers but in the electrical grid, the tariff schedules, the attendance records of engineering classrooms, and the willingness of a nation to keep building after the emergency that created its opportunity has faded.
Let me leave you with the forward-looking signals I am actually tracking. The short-term horizon, the next six months, belongs to transmission and permitting. I will be watching whether the CFE announces concrete grid investment plans with real numbers and real timelines. I will be watching for true announcements — not rumors — from Microsoft, Google, or Amazon about data center site selection in northern Mexico. And I will be watching the Commerce Department's export-control list with the anxiety of a trader watching an interest-rate decision. Each of those signals will tell us whether the current story is just narrative or whether the concrete is being poured.
The six-to-eighteen-month horizon is about rents and capital markets. I want to see the actual rental growth data for Monterrey and Chihuahua industrial properties, not the press-release versions. I want to see the filings of Mexican industrial REITs and infrastructure companies — the audited numbers, not the “AI narrative” slide decks. I want to see whether the kinds of institutional investors who allocate serious money to infrastructure are increasing their Mexican exposure. And I will be listening to the earnings calls of American hyperscalers for every mention of Mexico — measuring frequency as a proxy for seriousness.
The eighteen-to-thirty-six-month horizon is about electrical capacity and software sovereignty. I want to see the actual transfer capacity of the cross-border transmission lines, not just their announced plans. I want to see whether Mexico develops its own regulatory framework for data center operations, energy procurement, and data privacy — because regulatory maturity is a signal of long-term relevance. And I want to see whether Chinese hardware suppliers, through whatever structures they find lawful, deepen their presence in the Mexican market. Each of these will be a chapter in the larger story of whether the North American compute block becomes real.
One more thing. I am an analyst, and analysts are trained to hedge every claim. Let me hedge this one. My confidence in the Mexican infrastructure boom rests on assumptions that could break tomorrow: continued US capital expenditure on AI, unaffected by any downturn; continued political alignment between the two countries; continued American aversion to Chinese supply chains. Any one of those assumptions failing would trigger a cascade that would reach far beyond Mexican industrial parks. This is not a forecast of doom; it is a reminder of contingency. In the world of narrative analysis, the most dangerous mistake is to confuse the map with the territory.
Territory, in this case, is a strip of desert and industrial parks along a two-thousand-mile border. It is a place where American geopolitical anxiety meets Mexican economic pragmatism. It is a room full of concrete and copper and current and all the mundane engineering that makes the magic of artificial intelligence possible. The AI era will be remembered for many things: for the models, the agents, the breakthroughs, and the disasters. But underneath all of it, an older story runs like a current through the ground. The story of who supplies the watts that make the weights spin.
Reading the room in a room of code, I keep returning to the same conclusion. Mexico has been handed an unusual opportunity: to be woven, with varying degrees of consent, into the computational backbone of the most powerful economy on Earth. Whether that opportunity becomes a renaissance or a rental contract depends on decisions that are being made now, in boardrooms and in the offices of regulators, in engineering classrooms and in the dry corridors of the CFE. The momentum is real. The direction is favorable. But the deepest structural force at work is not Mexico's rise. It is the relocation of the physical architecture of intelligence — and whoever controls that architecture will own the terms of the next economy.
That is not a declaration. That is a question wrapped in a map. And like every map, it is more true about the terrain it describes than she who draws it can ever verify. I don’t know if the map's paths will hold. But I know where I will be standing when they are tested: at the intersection of watts and weights, watching the room, reading its code.
No, I don't have the final answer. I never do. That is the nature of being a narrative hunter. The stories recede as you approach them, and the map redraws itself as you read it. But the signal is loud, and the signal is physical. Mexico is no longer a place where AI infrastructure merely could be built. It is a place where, in the next years, we will discover whether the next economy has a foundation made of concrete, or only of promises.
I'll be there with my terminal, my Python scripts, and my unsolved question about what happens when the emergency fades. The lights will still be on. The question is whose hands will be holding the switch.
Reading the room in a room of code.
That is where the story begins to be true.